Multi-Agentic RAG with Hugging Face Code Brokers | by Gabriele Sgroi, PhD | Dec, 2024

Utilizing Qwen2.5–7B-Instruct powered code brokers to create a neighborhood, open supply, multi-agentic RAG system

Massive Language Fashions have proven spectacular capabilities and they’re nonetheless present process regular enhancements with every new era of fashions launched. Functions equivalent to chatbots and summarisation can straight exploit the language proficiency of LLMs as they’re solely required to provide textual outputs, which is their pure setting. Massive Language Fashions have additionally proven spectacular skills to know and remedy complicated duties, however so long as their answer stays “on paper”, i.e. in pure textual kind, they want an exterior consumer to behave on their behalf and report again the outcomes of the proposed actions. Agent programs remedy this downside by letting the fashions act on their surroundings, normally by way of a set of instruments that may carry out particular operations. On this approach, an LLM can discover options iteratively by trial and error whereas interacting with the surroundings.

An attention-grabbing scenario is when the instruments that an LLM agent has entry to are brokers themselves: that is the core idea of multi-agentic programs. A multi-agentic system solves duties by distributing and delegating duties to specialised fashions and placing their output collectively like puzzle items. A typical solution to implement such programs is by utilizing a supervisor agent to orchestrate and coordinate different brokers’ workflow.

Agentic programs, and particularly multi-agentic programs, require a strong LLM as a spine to carry out correctly, because the underlying mannequin wants to have the ability to perceive the aim and applicability of the assorted instruments in addition to break up the unique downside into sub-problems that may be tackled by every device. For that reason, proprietary fashions like ChatGpt or Anthropic’s Claude are typically the default go-to answer for agentic programs. Thankfully, open-source LLMs have continued to see enormous enhancements in efficiency a lot in order that a few of them now rival proprietary fashions in some situations. Much more curiously, modestly-sized open LLMs can now carry out complicated duties that have been unthinkable a few years in the past.

On this weblog publish, I’ll present how a “small” LLM that may run on shopper {hardware} is succesful sufficient to energy a multi-agentic system with good outcomes. Specifically, I’ll give a tutorial on how you should use Qwen2.5–7B-Instruct to create a multi-agentic RAG system. You’ll find the code implementation within the following GitHub repo and an illustrative Colab pocket book.

Earlier than diving into the main points of the system structure, I’ll recall some primary notions concerning LLM brokers that will likely be helpful to higher perceive the framework.

ReAct, proposed in ReAct: Synergizing Reasoning and Appearing in Language Fashions, is a well-liked framework for constructing LLM brokers. The primary concept of the tactic is to include the effectiveness of Chain of Thought prompting into an agent framework. ReACT consists of interleaved reasoning and motion steps: the Massive Language Mannequin is prompted to supply a thought sequence earlier than emitting an motion. On this approach the mannequin can create dynamic reasoning traces to steer actions and replace the high-level plan whereas incorporating info coming from the interplay with the surroundings. This enables for an iterative and incremental strategy to fixing the given process. In observe, the workflow of a ReAct agent is made up of Thought, Motion, and Remark sequences: the mannequin produces reasoning for a normal plan and particular device utilization within the Thought step, then invokes the related device within the Motion step, and at last receives suggestions from the surroundings within the Remark.

Beneath is an instance of what the ReACT framework appears like.

Code brokers are a specific sort of LLM brokers that use executable Python code to work together with the surroundings. They’re based mostly on the CodeAct framework proposed within the paper Executable Code Actions Elicit Higher LLM Brokers. CodeAct is similar to the ReAct framework, with the distinction that every motion consists of arbitrary executable code that may carry out a number of operations. Hand-crafted instruments are supplied to the agent as common Python features that it will probably name within the code.

Code brokers include a singular set of benefits over extra conventional brokers utilizing JSON or different textual content codecs to carry out actions:

• They will leverage present software program packages together with hand-crafted task-specific instruments.
• They will self-debug the generated code by utilizing the error messages returned after an error is raised.
• LLMs are conversant in writing code as it’s typically extensively current of their pre-training knowledge, making it a extra pure format to jot down their actions.
• Code naturally permits for the storage of intermediate outcomes and the composition of a number of operations in a single motion, whereas JSON or different textual content codecs may have a number of actions to perform the identical.

For these causes, Code Brokers can provide improved efficiency and quicker execution velocity than brokers utilizing JSON or different textual content codecs to execute actions.

Beneath is a concrete instance from the unique paper that showcases how code brokers can require fewer actions to unravel sure duties.

The Hugging Face transformers library gives helpful modules to construct brokers and, particularly, code brokers. The Hugging Face transformer brokers framework focuses on readability and modularity as core design rules. These are significantly essential when constructing an agent system: the complexity of the workflow makes it important to have management over all of the interconnected components of the structure. These design decisions make Hugging Face brokers an important device for constructing customized and versatile agent programs. When utilizing open-source fashions to energy the agent engine, the Hugging Face brokers framework has the additional benefit of permitting easy accessibility to the fashions and utilities current within the Hugging Face ecosystem.

Hugging Face code brokers additionally sort out the problem of insecure code execution. In actual fact, letting an LLM generate code unrestrained can pose critical dangers because it might carry out undesired actions. For instance, a hallucination might trigger the agent to erase essential recordsdata. As a way to mitigate this threat, Hugging Face code brokers implementation makes use of a ground-up strategy to safe code execution: the code interpreter can solely execute explicitly approved operations. That is in distinction to the standard top-down paradigm that begins with a totally purposeful Python interpreter after which forbids actions which may be harmful. The Hugging Face implementation features a record of secure, approved features that may be executed and gives a listing of secure modules that may be imported. The rest isn’t executable except it has been preemptively approved by the consumer. You possibly can learn extra about Hugging Face (code) brokers of their weblog posts:

Retrieval Augmented Technology has develop into the de facto customary for info retrieval duties involving Massive Language Fashions. It may possibly assist maintain the LLM info updated, give entry to particular info, and scale back hallucinations. It may possibly additionally improve human interpretability and supervision by returning the sources the mannequin used to generate its reply. The standard RAG workflow, consisting of a retrieval course of based mostly on semantic similarity to a consumer’s question and a mannequin’s context enhancement with the retrieved info, isn’t enough to unravel some particular duties. Some conditions that aren’t fitted to conventional RAG embody duties that want interactions with the data sources, queries needing a number of items of knowledge to be answered, and sophisticated queries requiring non-trivial manipulation to be related with the precise info contained within the sources.

A concrete difficult instance for conventional RAG programs is multi-hop query answering (MHQA). It entails extracting and mixing a number of items of knowledge, probably requiring a number of iterative reasoning processes over the extracted info and what’s nonetheless lacking. As an example, if the mannequin has been requested the query “Does birch plywood float in ethanol?”, even when the sources used for RAG contained details about the density of each supplies, the usual RAG framework might fail if these two items of knowledge usually are not straight linked.

A preferred solution to improve RAG to keep away from the abovementioned shortcomings is to make use of agentic programs. An LLM agent can break down the unique question right into a collection of sub-queries after which use semantic search as a device to retrieve passages for these generated sub-queries, altering and adjusting its plan as extra info is collected. It may possibly autonomously determine whether or not it has collected sufficient info to reply every question or if it ought to proceed the search. The agentic RAG framework may be additional enhanced by extending it to a multi-agentic system by which every agent has its personal outlined duties and duties. This enables, for instance, the separation between the high-level process planning and the interplay with the doc sources. Within the subsequent part, I’ll describe a sensible implementation of such a system.

On this part, I’ll talk about the overall architectural decisions I used to implement a Multi-Agentic RAG system based mostly on code brokers following the ReAct framework. You’ll find the remaining particulars within the full code implementation within the following GitHub repo.

The purpose of the multi-agentic system is to reply a query by looking the mandatory info on Wikipedia. It’s made up of three brokers:

• A supervisor agent whose job is to interrupt down the duty into sub-tasks and use their output to supply a closing reply.
• A Wikipedia search agent that finds related pages on Wikipedia and combines the data extracted from them.
• A web page search agent to retrieve and summarize info related to a given question from the supplied Wikipedia web page.

These three brokers are organized in a hierarchical style: every agent can use the agent instantly under within the hierarchy as a device. Specifically, the supervisor agent can name the Wikipedia search agent to seek out details about a question which, in flip, can use the web page search agent to extract specific info from Wikipedia pages.

Beneath is the diagram of the structure, specifying which hand-crafted instruments (together with instruments wrapping different brokers) every agent can name. Discover that since code brokers act utilizing code execution, these usually are not really the one instruments they’ll use as any native Python operation and performance (so long as it’s approved) can be utilized as properly.

Let’s dive into the main points of the workings of the brokers concerned within the structure.

Supervisor agent

That is the top-level agent, it receives the consumer’s query and it’s tasked to return a solution. It may possibly use the Wikipedia search agent as a device by prompting it with a question and receiving the ultimate outcomes of the search. Its objective is to gather the mandatory items of knowledge from Wikipedia by dividing the consumer query right into a collection of sub-queries and placing collectively the results of the search.

Beneath is the system immediate used for this agent. It’s constructed upon the default Hugging Face default immediate template. Discover that the examples supplied within the immediate observe the chat template of the mannequin powering the agent, on this case, Qwen2.5–7B-Instruct.

You might be an knowledgeable assistant who can discover reply on the web utilizing code blobs and instruments. To take action, you have got been given entry to a listing of instruments: these instruments are mainly Python features which you'll be able to name with code.
You may be given the duty of answering a consumer query and you need to reply it by retrieving the mandatory info from Wikipedia. Use and belief solely the data you retrieved, do not make up false information.
That can assist you, you have got been given entry to a search agent you should use as a device. You should use the search agent to seek out info on Wikipedia. Break down the duty into smaller sub-tasks and use the search agent to seek out the mandatory info for every sub-task.
To unravel the duty, you could plan ahead to proceed in a collection of steps, in a cycle of 'Thought:', 'Code:', and 'Remark:' sequences.
At every step, within the 'Thought:' sequence, you need to first clarify your reasoning in direction of fixing the duty and the instruments that you just need to use.
Then within the 'Code:' sequence, you need to write the code in easy Python. The code sequence should finish with '' sequence.
Throughout every intermediate step, you should use 'print()' to save lots of no matter essential info you'll then want. These print outputs will likely be supplied again to you by the consumer within the 'Remark:' discipline, which will likely be obtainable as enter for the following steps. At all times print the output of instruments, do not course of it or attempt to extract info earlier than inspecting it.
If an error rise whereas executing the code, will probably be proven within the 'Remark:' discipline. In that case, repair the code and take a look at once more.
Ultimately you must return a closing reply utilizing the `final_answer` device.
Listed here are a couple of notional examples:
---
<|im_start|>consumer
Job: When was the capital of Italy based?<|im_end|>
<|im_start|>assistant
Thought: Let's break up the duty: I first want to seek out the capital of Italy after which take a look at its basis date. I'll use the device `wikipedia_search_agent` to get the capital of Italy. Code:
```py
end result = wikipedia_search_agent("Italy capital")
print("Capital of Italy:", end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 0] -> Remark:
Capital of Italy:In line with the data extracted from the Wikipedia web page 'Rome', the capital of Italy is Rome.<|im_end|>
<|im_start|>assistant
Thought: Now that I do know that the capital of Italy is Rome, I can use the `wikipedia_search_agent` device to search for its basis date.
Code:
```py
end result = wikipedia_search_agent("Rome basis date")
print("Rome basis:", end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 1] -> Remark:
Rome basis: In line with the data from the Wikipedia web page 'Natale di Roma', the standard basis date of Rome is April 21, 753 BC.<|im_end|>
<|im_start|>assistant
Thought: Now that I've retrieved the related info, I can use the `final_answer` device to return the reply.
Code:
```py
final_answer("In line with the legend Rome was based on 21 April 753 BCE, however archaeological proof dates again its growth throughout the Bronze Age.")
```<|im_end|>
---
<|im_start|>consumer
Job: "What is the distinction in inhabitants between Shanghai and New York?"<|im_end|>
<|im_start|>assistant
Thought: I must get the populations for each cities and evaluate them: I'll use the device `search_agent` to get the inhabitants of each cities.
Code:
```py
population_guangzhou_info = wikipedia_search_agent("New York Metropolis inhabitants")
population_shanghai_info = wikipedia_search_agent("Shanghai inhabitants")
print("Inhabitants Guangzhou:", population_guangzhou)
print("Inhabitants Shanghai:", population_shanghai)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 0] -> Remark:
Inhabitants Guangzhou: The inhabitants of New York Metropolis is roughly 8,258,035 as of 2023.
Inhabitants Shanghai: In line with the data extracted from the Wikipedia web page 'Shanghai', the inhabitants of the town correct is round 24.87 million inhabitants in 2023.<|im_end|>
<|im_start|>assistant
Thought: Now I do know each the inhabitants of Shanghai (24.87 million) and of New York Metropolis (8.25 million), I'll calculate the distinction and return the end result.
Code:
```py
population_difference = 24.87*1e6 - 8.25*1e6
reply=f"The distinction in inhabitants between Shanghai and New York is {population_difference} inhabitants."
final_answer(reply)
```<|im_end|>
---
On high of performing computations within the Python code snippets that you just create, you have got entry to these instruments (and no different device):
<>
<>
You should use imports in your code, however completely from the next record of modules: <>.  Don't attempt to import different modules or else you'll get an error.
Now begin and remedy the duty!

Wikipedia search agent

This agent stories to the supervisor agent, it receives a question from it and it’s tasked to return the data it has retrieved from Wikipedia. It may possibly entry two instruments:

• A Wikipedia search device, utilizing the built-in search operate from the wikipedia bundle. It receives a question as enter and returns a listing of Wikipedia pages and their summaries.
• A web page search agent that retrieves details about a question from a particular Wikipedia web page.

This agent collects the data to reply the question, dividing it into additional sub-queries, and mixing info from a number of pages if wanted. That is completed by utilizing the search device of the wikipedia bundle to establish potential pages that may include the mandatory info to reply the question: the agent can both use the reported web page summaries or name the web page search agent to extract extra info from a particular web page. After sufficient knowledge has been collected, it returns a solution to the supervisor agent.

The system immediate is once more a slight modification of the Hugging Face default immediate with some particular examples following the mannequin’s chat template.

You might be an knowledgeable assistant that retrieves info from Wikipedia utilizing code blobs and instruments. To take action, you have got been given entry to a listing of instruments: these instruments are mainly Python features which you'll be able to name with code.
You may be given a normal question, your process will likely be of retrieving and summarising info that's related to the question from a number of passages retrieved from the given Wikipedia web page. Use and belief solely the data you retrieved, do not make up false information. Attempt to summarize the data in a couple of sentences.
To unravel the duty, you could plan ahead to proceed in a collection of steps, in a cycle of 'Thought:', 'Code:', and 'Remark:' sequences.
At every step, within the 'Thought:' sequence, you need to first clarify your reasoning in direction of fixing the duty and the instruments that you just need to use.
Then within the 'Code:' sequence, you need to write the code in easy Python. The code sequence should finish with '' sequence.
Throughout every intermediate step, you should use 'print()' to save lots of no matter essential info you'll then want. These print outputs will likely be supplied again to you by the consumer within the 'Remark:' discipline, which will likely be obtainable as enter for the following steps. At all times print the output of instruments, do not course of it or attempt to extract info earlier than inspecting it.
If an error rise whereas executing the code, will probably be proven within the 'Remark:' discipline. In that case, repair the code and take a look at once more.
Ultimately you must return a closing reply utilizing the `final_answer` device.
Listed here are a couple of notional examples:
---
<|im_start|>consumer
Job: Retrieve details about the question:"What is the capital of France?" from the Wikipedia web page "France".<|im_end|>
<|im_start|>assistant
Thought: I would like to seek out the capital of France. I'll use the device `retrieve_passages` to get the capital of France from the Wikipedia web page.
Code:
```py
end result = retrieve_passages("France capital")
print("Capital of France:", end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 0] -> Remark:
Retrieved passages for question "France capital":
Passage 0: ... inhabitants of almost 68.4 million as of January 2024. France is a semi-presidential republic with its capital in Paris, the ...
Passage 1: ... France, formally the French Republic, is a rustic positioned primarily in Western Europe. Its abroad areas and territories ...
Passage 2: ... The overwhelming majority of France's territory and inhabitants is located in Western Europe and is named Metropolitan France. It's ...
Passage 3: ... France is a extremely urbanised nation, with its largest cities (by way of metropolitan space inhabitants in 2021) being Paris ...
Passage 4: ... === Authorities ===nFrance.fr – official French tourism website (in English)...<|im_end|>
<|im_start|>assistant
Thought: Now that I do know that the capital of France is Paris, I can use the `final_answer` device to return the reply.
Code:
```py
final_answer("The capital of France is Paris.")
```<|im_end|>
---
<|im_start|>consumer
Job: Retrieve details about the question:"Tallest mountain within the World" from the Wikipedia web page "Listing of highest mountains on Earth"<|im_end|>
<|im_start|>assistant
Thought: I would like to seek out the tallest mountain on the earth. I'll use the device `retrieve_passages` to search for knowledge on the Wikipedia web page.
Code:
```py
end result = retrieve_passages("highest mountain")
print(end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 1] -> Remark:
Retrieved passages for question "highest mountain":
Passage 0: ... above sea degree) is the world's tallest mountain and volcano, rising about 10,203 m (33,474 ft) from the Pacific Ocean flooring. ...
Passage 1: ... As of December 2018, the very best peaks on 4 of the mountains—Gangkhar Puensum, Labuche Kang III, Karjiang, and Tongshanjiabu, all positioned in Bhutan or China—haven't been ascended. ...
Passage 2: ... The very best mountains above sea degree are typically not the very best above the encircling terrain. ...
Passage 3: ... The very best mountain outdoors of Asia is Aconcagua (6,961 m or 22,838 ft), the 189th highest on the earth. ...
Passage 4: ... the southern summit of Peru's tallest mountain, Huascarán, is one other contender. Each have elevations above sea degree greater than 2 km (1.2 mi) lower than that of Everest....
<|im_end|>
<|im_start|>assistant
Thought: The outcomes do not clearly specify a transparent end result for the world's tallest mountain, I'll use the device `web_results` with a unique question.
Code:
```py
end result = retrieve_passages("world's tallest mountain")
print(end result)
```<|im_end|>
<|im_start|>consumer
Passages retrieved from web page Listing of highest mountains on Earth:
Passage 0: ... The very best mountain outdoors of Asia is Aconcagua (6,961 m or 22,838 ft), the 189th highest on the earth....
Passage 1: ... above sea degree) is the world's tallest mountain and volcano, rising about 10,203 m (33,474 ft) from the Pacific Ocean flooring. ...
Passage 2: ... The bases of mountain islands are under sea degree, and given this consideration Mauna Kea (4,207 m (13,802 ft) above sea degree) is the world's tallest mountain and volcano, rising about 10,203 m (33,474 ft) from the Pacific Ocean flooring. ...
Passage 3: ... the southern summit of Peru's tallest mountain, Huascarán, is one other contender. Each have elevations above sea degree greater than 2 km (1.2 mi) lower than that of Everest. ...
Passage 4: ... The very best mountains are additionally not typically essentially the most voluminous. Mauna Loa (4,169 m or 13,678 ft) is the biggest mountain on Earth by way of base space (about 5,200 km2 or 2,000 sq mi) and quantity (about 42,000 km3 or 10,000 cu mi)...<|im_end|>
<|im_start|>assistant
Thought: I've discovered that Mauna Kea is the world's tallest mountain rising about 10,203 m (33,474 ft) from the Pacific Ocean flooring. I can use the `final_answer` device to return the related info.
Code:
```py
final_answer("Mauna Kea is the world's tallest mountain, rising about 10,203 m (33,474 ft) from the Pacific Ocean flooring.")
```<|im_end|>
___
On high of performing computations within the Python code snippets that you just create, you have got entry to these instruments (and no different device):
<>
<>
You should use imports in your code, however solely from the next record of modules: <>. Don't attempt to import different modules or else you'll get an error.
Now begin and remedy the duty!

Web page search agent

This agent stories to the Wikipedia search agent, which gives it with a question and the title of a Wikipedia web page, and it’s tasked to retrieve the related info to reply the question from that web page. That is, in essence, a single-agent RAG system. To carry out the duty, this agent generates customized queries and makes use of the semantic search device to retrieve the passages which might be extra much like them. The semantic search device follows a easy implementation that splits the web page contents into chunks and embeds them utilizing the FAISS vector database supplied by LangChain.

Beneath is the system immediate, nonetheless constructed upon the one supplied by default by Hugging Face

You might be an knowledgeable assistant that finds solutions to questions by consulting Wikipedia, utilizing code blobs and instruments. To take action, you have got been given entry to a listing of instruments: these instruments are mainly Python features which you'll be able to name with code.
You may be given a normal question, your process will likely be of discovering a solution to the question utilizing the data you retrieve from Wikipedia. Use and belief solely the data you retrieved, do not make up false information. Cite the web page the place you discovered the data.
You possibly can seek for pages and their summaries from Wikipedia utilizing the `search_wikipedia` device and search for particular passages from a web page utilizing the `search_info` device. It's best to determine the best way to use these instruments to seek out an acceptable reply:some queries may be answered by  one web page abstract, others can require  particular passages from a number of pages.
To unravel the duty, you could plan ahead to proceed in a collection of steps, in a cycle of 'Thought:', 'Code:', and 'Remark:' sequences.
At every step, within the 'Thought:' sequence, you need to first clarify your reasoning in direction of fixing the duty and the instruments that you just need to use.
Then within the 'Code:' sequence, you need to write the code in easy Python. The code sequence should finish with '' sequence.
Throughout every intermediate step, you should use 'print()' to save lots of no matter essential info you'll then want. These print outputs will likely be supplied again to you by the consumer within the 'Remark:' discipline, which will likely be obtainable as enter for the following steps. At all times print the output of instruments, do not course of it or attempt to extract info earlier than inspecting it.
If an error rise whereas executing the code, will probably be proven within the 'Remark:' discipline. In that case, repair the code and take a look at once more.
Ultimately you must return a closing reply utilizing the `final_answer` device.
Listed here are a couple of notional examples:
---
<|im_start|>consumer
Job: When was the traditional thinker Seneca born?<|im_end|>
<|im_start|>assistant
Thought: I'll use the device `search_wikipedia` to seek for Seneca's start on Wikipedia. I'll specify I'm on the lookout for the thinker for disambiguation.
Code:
```py
end result = search_wikipedia("Seneca thinker start")
print("end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 0] -> Remark:
Pages discovered for question 'Seneca thinker start':
Web page: Seneca the Youthful
Abstract: Lucius Annaeus Seneca the Youthful ( SEN-ik-ə; c.4 BC – AD 65), normally identified mononymously as Seneca, was a Stoic thinker of Historic Rome, a statesman, dramatist, and in a single work, satirist, from the post-Augustan age of Latin literature.
Seneca was born in Colonia Patricia Corduba in Hispania, a
Web page: Phaedra (Seneca)
Abstract: Phaedra is a Roman tragedy written by thinker and dramatist Lucius Annaeus Seneca earlier than 54 A.D. Its 1,280 traces of verse inform the story of Phaedra, spouse of King Theseus of Athens and her consuming lust for her stepson Hippolytus. Primarily based on Greek mythology and the tragedy Hippolytus by Euripides,
Web page: Seneca the Elder
Abstract: Lucius Annaeus Seneca the Elder ( SEN-ik-ə; c.54 BC – c. AD 39), also referred to as Seneca the Rhetorician, was a Roman author, born of a rich equestrian household of Corduba, Hispania. He wrote a set of reminiscences concerning the Roman colleges of rhetoric, six books of that are extant in a extra or
Web page: AD 1
Abstract: AD 1 (I) or 1 CE was a typical 12 months beginning on Saturday or Sunday, a typical 12 months beginning on Saturday by the proleptic Julian calendar, and a typical 12 months beginning on Monday by the proleptic Gregorian calendar. It's the epoch 12 months for the Anno Domini (AD) Christian calendar period, and the first 12 months of
Web page: Seneca Falls Conference
Abstract: The Seneca Falls Conference was the primary ladies's rights conference. It marketed itself as "a conference to debate the social, civil, and non secular situation and rights of girl". Held within the Wesleyan Chapel of the city of Seneca Falls, New York, it spanned two days over July 19–20, 1848.  Attrac
<|im_start|>assistant
Thought: From the abstract of the web page "", I can see that Seneca was born in . I can use the `final_answer` device to return the reply.
Code:
```py
final_answer("In line with the Wikipedia web page 'Seneca the Youthful', Seneca was born in 4 BC.")
```<|im_end|>
---
<|im_start|>consumer
Job: Who was Charlemagne predecessor?<|im_end|>
<|im_start|>assistant
Thought: I'll use the device `search_wikipedia` to seek for Charlemagne reign period.
Code:
```py
end result = search_wikipedia("Charlemagne predecessor")
print(end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 0] -> Remark:
Pages discovered for question 'Charlemagne predecessor':
Web page: Charlemagne
Abstract: Charlemagne ( SHAR-lə-mayn; 2 April 748 – 28 January 814) was King of the Franks from 768, King of the Lombards from 774, and Emperor of what's now generally known as the Carolingian Empire from 800, holding these titles till his dying in 814. He united most of Western and Central Europe, and was the primary
Web page: Pope Leo III
Abstract: Pope Leo III (Latin: Leo III; died 12 June 816) was bishop of Rome and ruler of the Papal States from 26 December 795 to his dying. Protected by Charlemagne from the supporters of his predecessor, Adrian I, Leo subsequently strengthened Charlemagne's place by crowning him emperor. The coronation
Web page: Throne of Charlemagne
Abstract: The Throne of Charlemagne (German: Karlsthron or Aachener Königsthron, "Royal Throne of Aachen") is a throne erected within the 790s by Charlemagne, as one of many fittings of his palatine chapel in Aachen (right now's Aachen Cathedral) and positioned within the Octagon of the church. Till 1531, it served because the co
Web page: Louis the Pious
Abstract: Louis the Pious (Latin: Hludowicus Pius; French: Louis le Pieux; German: Ludwig der Fromme; 16 April 778 – 20 June 840), additionally referred to as the Truthful and the Debonaire, was King of the Franks and co-emperor along with his father, Charlemagne, from 813. He was additionally King of Aquitaine from 781. As the one surviving
Web page: Holy Roman Emperor
Abstract: The Holy Roman Emperor, initially and formally the Emperor of the Romans (Latin: Imperator Romanorum; German: Kaiser der Römer) throughout the Center Ages, and also referred to as the Romano-German Emperor for the reason that early trendy interval (Latin: Imperator Germanorum; German: Römisch-deutscher Kaiser, lit. 'R
<|im_end|>
<|im_start|>assistant
Thought: The outcomes do not include specific details about Charlemagne predecessor, I'll seek for extra info on the web page 'Charlemagne' utilizing the 'search_info' device.
Code:
```py
end result = search_info("Charlemagne predecessor", "Charlemagne")
print(end result)
```<|im_end|>
<|im_start|>consumer
[OUTPUT OF STEP 1] -> Remark:
Info retrieved from the web page 'Charlemagne' for the question 'Charlemagne predecessor':
Charlemagne's predecessor was Pepin the Quick.
<|im_end|>
<|im_start|>assistant
Thought: I've discovered that, in response to the Wikipedia web page 'Charlemagne', Pepin the Quick was Charlemagne predecessor. I'll return the outcomes utilizing the `final_answer` device.
Code:
```py
final_answer("In line with the data extracted from the Wikipedia web page 'Charlemagne', his predecessor was Pepin the Quick.")
```<|im_end|>
___
On high of performing computations within the Python code snippets that you just create, you have got entry to these instruments (and no different device):
<>
<>
You should use imports in your code, however solely from the next record of modules: <>. Don't attempt to import different modules or else you'll get an error.
Now begin and remedy the duty!

Implementation decisions

On this subsection, I’ll define the details that differ from what may very well be a simple implementation of the structure utilizing Hugging Face brokers. These are the outcomes of restricted trial and error earlier than acquiring an answer that works moderately properly. I haven’t carried out intensive testing and ablations so that they might not be the optimum decisions.

• Prompting: as defined within the earlier sections, every agent has its personal specialised system immediate that differs from the default one supplied by Hugging Face Code Brokers. I noticed that, maybe as a result of restricted measurement of the mannequin used, the overall customary system immediate was not giving good outcomes. The mannequin appears to work greatest with a system immediate that displays intently the duties it’s requested to carry out, together with tailor-made examples of serious use instances. Since I used a chat mannequin with the purpose of bettering instruction following conduct, the supplied examples observe the mannequin’s chat template to be as shut as attainable to the format encountered throughout a run.
• Summarizing historical past: lengthy execution histories have detrimental results on each execution velocity and process efficiency. The latter may very well be as a result of restricted capability of the mannequin to retrieve the mandatory info from an extended context. Furthermore, extraordinarily lengthy execution histories might exceed the utmost context size for the engine mannequin. To mitigate these issues and velocity up execution, I selected to not present all the main points of the earlier thought-action-observation steps, however as an alternative collected solely the earlier observations. Extra particularly, at every step the mannequin solely receives the next chat historical past: the system message, the primary message containing the duty, its final motion, and all of the historical past of the earlier observations. Moreover, execution errors are current within the commentary historical past provided that they occur within the final step, earlier errors which have been already solved are discarded.
• Instruments vs managed brokers: Hugging Face brokers implementation has native help for managed brokers however wrapping them as instruments permits for higher management of the prompts and a extra streamlined implementation. Specifically, Hugging Face implementation provides specific prompts to each the managed brokers and their managers. Whereas I haven’t seen substantial variations within the capability to unravel the given process, I most popular the second strategy as it’s extra versatile and streamlined for the offered structure, and it permits for simpler management over the brokers’ conduct. This additionally helps with lowering the immediate size, which is helpful for dashing up computations.
• Restrict the utmost variety of trials for the web page search agent: typically the web page search agent retains on the lookout for info on a given web page that doesn’t include it. Lowering the utmost variety of trials mitigated this concern: after reaching that quantity the agent execution is stopped and the device returns the final commentary from code execution.
• Altering device response to consumer message: that is extra of a technical, implementation-specific level. Because the solely supported roles for the chat template of Qwen2.5–7B-Instruct are system, consumer, and assistant, observations are returned as consumer messages.

On this part, I’ll gather some examples produced with the described structure. I may also present the detailed execution of all of the brokers concerned within the process to showcase how they work together in some actual instances. You possibly can discover extra examples utilizing the illustrative Colab pocket book. All of the extracted passages are taken from the Wikipedia web page specified within the execution.

Let’s begin exploring a easy instance that requires combining info from a number of sources and a few primary arithmetic understanding.

Job: What number of years handed between the development of the Eiffel Tower and the development of the Empire State Constructing?

Beneath are the logs of the execution of the supervisor agent. It may possibly appropriately break up the duty into discovering the development durations for each buildings and compute the 12 months distinction.

Agent:  Thought: First, I would like to seek out out when the Eiffel Tower and the Empire State Constructing have been constructed. I'll use the `wikipedia_search_agent` device to seek out this info.
Code:
```py
wikipedia_search_agent(question="Eiffel Tower development 12 months")
```
Remark:  Print outputs:
Final output from code snippet:
In line with the data extracted from the Wikipedia web page 'Eiffel Tower', the development of the Eiffel Tower occurred from 1887 to 1889.
Agent:  Thought: Now I've the development interval of the Eiffel Tower. Subsequent, I would like to seek out the development interval of the Empire State Constructing. I'll use the `wikipedia_search_agent` device once more.
Code:
```py
wikipedia_search_agent(question="Empire State Constructing development 12 months")
```
Remark:  Print outputs:
Final output from code snippet:
In line with the data extracted from the Wikipedia web page 'Empire State Constructing', the development of the Empire State Constructing began in 1930 and resulted in 1931.
Agent:  Thought: Now I've the development durations for each the Eiffel Tower and the Empire State Constructing. I can now calculate the variety of years between their constructions. 
Code:
```py
construction_eiffel_tower_start = 1887
construction_eiffel_tower_end = 1889
construction_empire_state_building_start = 1930
construction_empire_state_building_end = 1931
years_passed = construction_empire_state_building_start - construction_eiffel_tower_end
reply = f"The variety of years handed between the development of the Eiffel Tower and the Empire State Constructing is {years_passed} years."
final_answer(reply)
```
Remark:  Print outputs:
Final output from code snippet:
The variety of years handed between the development of the Eiffel Tower and the Empire State Constructing is 41 years.

Let’s now visualize the execution of the Wikipedia search agent. It appropriately searches and stories the development durations discovered within the abstract of the pages, with no need to examine the person pages. It’s attention-grabbing to notice that regardless of being requested generically concerning the “development 12 months”, it stories the complete development interval as it isn’t clear if the 12 months refers back to the begin or the tip of the development works.

TASK:  Eiffel Tower development 12 months
AGENT:  Thought: I'll use the `search_wikipedia` device to seek out details about the Eiffel Tower development 12 months.
Code:
```py
search_wikipedia('Eiffel Tower development 12 months')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Pages discovered for question 'Eiffel Tower development 12 months':
Web page: Eiffel Tower
Abstract: The Eiffel Tower (  EYE-fəl; French: Tour Eiffel [tuʁ ɛfɛl] ) is a wrought-iron lattice tower on the Champ de Mars in Paris, France. It's named after the engineer Gustave Eiffel, whose firm designed and constructed the tower from 1887 to 1889.
Regionally nicknamed "La dame de fer" (French for "Iron Girl"), it was constructed because the centerpiece of the 1889 World's Truthful, and to crown the centennial anniversary of the French Revolution. Though initially criticised by a few of France's main artists and intellectuals for its design, it has since develop into a world cultural icon of France and one of the vital recognisable buildings on the earth. The tower acquired 5,889,000 guests in 2022. The Eiffel Tower is essentially the most visited monument with an entrance charge on the earth: 6.91 million folks ascended it in 2015. It was designated a monument historique in 1964, and was named a part of a UNESCO World Heritage Web site ("Paris, Banks of the Seine") in 1991.
The tower is 330 metres (1,083 ft) tall, about t
Web page: Eiffel Tower (Paris, Texas)
Abstract: Texas's Eiffel Tower is a landmark within the metropolis of Paris, Texas. The tower was constructed in 1993. It's a scale mannequin of the Eiffel Tower in Paris, France; at 65 toes in peak, it's roughly one-sixteenth of the peak of the unique. 
Web page: Gustave Eiffel
Abstract: Alexandre Gustave Eiffel (  EYE-fəl, French: [alɛksɑ̃dʁ ɡystav ɛfɛl]; né Bonickhausen dit Eiffel; 15 December 1832 – 27 December 1923) was a French civil engineer. A graduate of École Centrale des Arts et Manufactures, he made his title with numerous bridges for the French railway community, most famously the Garabit Viaduct. He's greatest identified for the world-famous Eiffel Tower, designed by his firm and constructed for the 1889 Common Exposition in Paris, and his contribution to constructing the Statue of Liberty in New York. After his retirement from engineering, Eiffel targeted on analysis into meteorology and aerodynamics, making vital contributions in each fields.
Web page: Watkin's Tower
Abstract: Watkin's Tower was {a partially} accomplished iron lattice tower in Wembley Park, London, England. Its development was an formidable mission to create a 358-metre (1,175 ft)-high customer attraction in Wembley Park to the north of the town, led by the railway entrepreneur Sir Edward Watkin. Marketed because the "Nice Tower of London", it was designed to surpass the peak of the Eiffel Tower in Paris, and it was a part of Wembley Park's emergence as a leisure place. The tower was by no means accomplished and it was demolished in 1907. The positioning of the tower is now occupied by the English nationwide soccer floor, Wembley Stadium.
Web page: Eiffel Tower (Paris, Tennessee)
Abstract: The Eiffel Tower is a landmark within the metropolis of Paris, Tennessee. It's a 1:20 scale duplicate of the unique positioned in Paris, France.
AGENT:  Thought: From the abstract of the web page 'Eiffel Tower', I can see that the development of the Eiffel Tower occurred from 1887 to 1889. I'll use the `final_answer` device to return the reply.
Code:
```py
final_answer("In line with the data extracted from the Wikipedia web page 'Eiffel Tower', the development of the Eiffel Tower occurred from 1887 to 1889.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
In line with the data extracted from the Wikipedia web page 'Eiffel Tower', the development of the Eiffel Tower occurred from 1887 to 1889.
------------------------------
TASK:  Empire State Constructing development 12 months
AGENT:  Thought: I'll use the `search_wikipedia` device to seek out details about the development 12 months of the Empire State Constructing.
Code:
```py
search_wikipedia('Empire State Constructing development 12 months')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Pages discovered for question 'Empire State Constructing development 12 months':
Web page: Empire State Constructing
Abstract: The Empire State Constructing is a 102-story Artwork Deco skyscraper within the Midtown South neighborhood of Manhattan in New York Metropolis. The constructing was designed by Shreve, Lamb & Harmon and constructed from 1930 to 1931. Its title is derived from "Empire State", the nickname of the state of New York. The constructing has a roof peak of 1,250 toes (380 m) and stands a complete of 1,454 toes (443.2 m) tall, together with its antenna. The Empire State Constructing was the world's tallest constructing till the primary tower of the World Commerce Middle was topped out in 1970; following the September 11 assaults in 2001, the Empire State Constructing was New York Metropolis's tallest constructing till it was surpassed in 2012 by One World Commerce Middle. As of 2024, the constructing is the seventh-tallest constructing in New York Metropolis, the ninth-tallest accomplished skyscraper in the USA, and the 57th-tallest accomplished skyscraper on the earth.
The positioning of the Empire State Constructing, on the west facet of Fifth Avenue between West thirty third and thirty fourth St
Web page: British Empire Constructing
Abstract: The British Empire Constructing, additionally identified by its tackle 620 Fifth Avenue, is a industrial constructing at Rockefeller Middle within the Midtown Manhattan neighborhood of New York Metropolis. Accomplished in 1933, the six-story construction was designed within the Artwork Deco model by Raymond Hood, Rockefeller Middle's lead architect. The British Empire Constructing, together with the almost equivalent La Maison Francaise to the south and the high-rise Worldwide Constructing to the north, comprise a gaggle of retail-and-office buildings generally known as the Worldwide Advanced. La Maison Francaise and the British Empire Constructing are separated by Channel Gardens, a planted pedestrian esplanade operating west to the complicated's Decrease Plaza.
The facade is fabricated from limestone, with a foremost entrance alongside Fifth Avenue and secondary entrances on fiftieth Avenue and Channel Gardens. The highest of the British Empire Constructing incorporates setbacks, a rooftop backyard, and a partial seventh-story penthouse. The constructing's entrances include ornate ornament
Web page: 2012 Empire State Constructing taking pictures
Abstract: On August 24, 2012, a gunman shot and killed a former co-worker outdoors the Empire State Constructing in New York Metropolis. Following the preliminary taking pictures, the gunman, 58-year-old Jeffrey T. Johnson, was fatally shot by cops after elevating his weapon at them. 9 bystanders have been wounded by stray bullets fired by the officers and ricocheting particles, however none suffered life-threatening accidents.
Web page: Empire State Constructing in well-liked tradition
Abstract: The Empire State Constructing is a 102-story skyscraper on Fifth Avenue between West thirty third and thirty fourth Streets in Midtown, Manhattan, New York Metropolis. It has a roof peak of 1,250 toes (381 m), and with its antenna included, it stands a complete of 1,454 toes (443.2 m) tall. The Empire State Constructing stood because the world's tallest constructing for almost 40 years, from its completion in early 1931 till the topping out of the unique World Commerce Middle's North Tower in late 1970. It has been featured in quite a few movies, TV exhibits, songs, video video games, books, images, and paintings.
Web page: Empire State Plaza
Abstract: The Governor Nelson A. Rockefeller Empire State Plaza (identified generally because the Empire State Plaza, and in addition because the South Mall) is a fancy of a number of state authorities buildings in downtown Albany, New York.
The complicated was constructed between 1965 and 1976 at an estimated whole price of $2 billion. It homes a number of departments of the New York State administration and is built-in with the New York State Capitol, accomplished in 1899, which homes the state legislature. Among the many workplaces on the plaza are the Division of Well being and the Biggs Laboratory of the Wadsworth Middle. The Empire State Artwork Assortment, a significant public assortment of Nineteen Sixties and Nineteen Seventies monumental summary artworks, is on everlasting show all through the location. The New York State Workplace of Basic Companies maintains the plaza. The Nelson A. Rockefeller Empire State Plaza Performing Arts Middle Company is a New York state public-benefit company that was created in 1979 to handle the performing arts facility within the plaza.
AGENT:  Thought: From the abstract of the web page "Empire State Constructing", I can see that the constructing was designed and constructed from 1930 to 1931. I'll use the `final_answer` device to return the reply.
Code:
```py
final_answer("In line with the data extracted from the Wikipedia web page 'Empire State Constructing', the development of the Empire State Constructing began in 1930 and resulted in 1931.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
In line with the data extracted from the Wikipedia web page 'Empire State Constructing', the development of the Empire State Constructing began in 1930 and resulted in 1931.

The passage retrieval agent was not used as all the data was already discovered within the summaries of the pages.

Let’s now strive a extra sophisticated instance asking if birch plywood floats in alcohol. This query requires primary physics understanding and arithmetic skills, mixed with non-trivial retrieval duties.

Job: Does birch plywood float in ethanol?

Let’s first visualize the logs of the assorted brokers.

Supervisor agent logs

AGENT:  Thought: To find out if birch plywood floats in ethanol, we have to know the density of each birch plywood and ethanol. If the density of birch plywood is lower than the density of ethanol, it would float. I'll use the `wikipedia_search_agent` device to seek out the densities of birch plywood and ethanol.
Code:
```py
birch_plywood_density_info = wikipedia_search_agent("density of birch plywood")
ethanol_density_info = wikipedia_search_agent("density of ethanol")
print(birch_plywood_density_info)
print(ethanol_density_info)
```
OBSERVATION:  Print outputs:
The density of birch plywood is roughly 680 kg/m³.
The density of ethanol at 20°C is roughly 0.7894 g/cm³.
AGENT:  Thought: Now that I've the densities, I can evaluate them. Birch plywood has a density of 680 kg/m³ and ethanol has a density of 0.7894 g/cm³. Since 1 g/cm³ is the same as 1000 kg/m³, the density of ethanol is 789.4 kg/m³. Because the density of birch plywood (680 kg/m³) is lower than the density of ethanol (789.4 kg/m³), birch plywood will float in ethanol. I'll now present this conclusion utilizing the `final_answer` operate.
Code:
```py
final_answer("Birch plywood will float in ethanol as a result of its density (680 kg/m³) is lower than the density of ethanol (789.4 kg/m³).")
```
OBSERVATION:  Print outputs:
Final output from code snippet:
Birch plywood will float in ethanol as a result of its density (680 kg/m³) is lower than the density of ethanol (789.4 kg/m³).

Wikipedia search agent logs

TASK:  density of birch plywood
AGENT:  Thought: I'll use the `search_wikipedia` device to seek out details about the density of birch plywood. 
Code:
```py
search_wikipedia('birch plywood')
```
OBSERVATION:  Print outputs:
Final output from code snippet:
Pages discovered for question 'birch plywood':
Web page: Plywood
Abstract: Plywood is a composite materials manufactured from skinny layers, or "plies", of wooden veneer which have been stacked and glued collectively. It's an engineered wooden from the household of manufactured boards, which embody plywood, medium-density fibreboard (MDF), oriented strand board (OSB), and particle board (or chipboard).
All plywoods bind resin and wooden fibre sheets (cellulose cells are lengthy, sturdy and skinny) to kind a composite materials. The sheets of wooden are stacked such that every layer has its grain set sometimes (see under) perpendicular to its adjoining layers. This alternation of the grain is named cross-graining and has a number of essential advantages: it reduces the tendency of wooden to separate when nailed on the edges; it reduces thickness swelling and shrinkage, offering improved dimensional stability; and it makes the power of the panel constant throughout all instructions.  There may be normally an odd variety of plies, in order that the sheet is balanced, that's, the floor layers have their gr
Web page: Birch
Abstract: A birch is a thin-leaved deciduous hardwood tree of the genus Betula (), within the household Betulaceae, which additionally consists of alders, hazels, and hornbeams. It's intently associated to the beech-oak household Fagaceae. The genus Betula incorporates 30 to 60 identified taxa of which 11 are on the IUCN 2011 Pink Listing of Threatened Species. They're sometimes short-lived pioneer species and are widespread within the Northern Hemisphere, significantly in northern areas of temperate climates and in boreal climates. Birch wooden is used for a variety of functions.
Web page: Birch wooden
Abstract: Birch wooden is a kind of wooden of the birch. Birch wooden is pale yellow-brown wooden having an in depth, straight grain and uniform texture that finishes to a clean floor. Typically it's dyed to mimic mahogany. The sort of wooden is used for amongst others firewood, turnery, furnishings, cabinetry, instruments handles, hoops, plywood, flooring and shoe heels.
Web page: Falco F8L
Abstract: The Falco F8L is an Italian-designed light-weight 2-seater aerobatic plane.
Web page: Crusader (speedboat)
Abstract: Crusader was a jet-powered velocity boat piloted by John Cobb.
The mix of an aerodynamically secure hull kind and turbojet propulsion was proposed by Reid Railton, Cobb's adviser. A rocket-powered scale mannequin was examined at Haslar. The total measurement design was by Peter du Cane and constructed by Vospers of Portsmouth. Technical help got here from Saunders-Roe and Vickers-Supermarine. It price £15,000 in 1949.
It was silver and scarlet in color and 10 m lengthy.  
The engine was a de Havilland Ghost Mk 48 centrifugal turbojet supplied as a mortgage by the Ministry of Provide on the request of Main Frank Halford, the engine designer. The engine was rated at 5,000 lb thrust fed by two scoop inlets ahead of the cockpit.
The hull was of trimaran kind, a foremost hull with a planing step, and two smaller rear-mounted outriggers. Development was of birch plywood frames and stringers. The hull was skinned in birch ply lined in doped material with metallic pores and skin reinforcement for planing surfaces. Plane-sty
AGENT:  Thought: The pages returned from the search didn't straight point out the density of birch plywood. Nevertheless, the web page on Birch Wooden mentions that birch wooden is used for plywood. I'll use the `search_info` device to seek out extra particular details about the density of birch plywood from the web page 'Plywood'.
Code:
```py
search_info('density of birch plywood', 'Plywood')
```
OBSERVATION:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Plywood' for the question 'density of birch plywood':
The density of birch plywood is roughly 680 kg/m³.
AGENT:  Thought: I've discovered the density of birch plywood from the Wikipedia web page 'Plywood'. The density is roughly 680 kg/m³.
Code:
```py
final_answer("The density of birch plywood is roughly 680 kg/m³.")
```
OBSERVATION:  Print outputs:
Final output from code snippet:
The density of birch plywood is roughly 680 kg/m³.
------------------------------
TASK:  density of ethanol
AGENT:  Thought: I'll use the `search_wikipedia` device to seek out the density of ethanol from Wikipedia. 
Code:
```py
search_wikipedia('density of ethanol')
```
OBSERVATION:  Print outputs:
Final output from code snippet:
Pages discovered for question 'density of ethanol':
Web page: Ethanol (knowledge web page)
Abstract: This web page gives supplementary chemical knowledge on ethanol.
Web page: Ethanol
Abstract: Ethanol (additionally referred to as ethyl alcohol, grain alcohol, ingesting alcohol, or just alcohol) is an natural compound with the chemical system CH3CH2OH. It's an alcohol, with its system additionally written as C2H5OH, C2H6O or EtOH, the place Et stands for ethyl. Ethanol is a risky, flammable, colorless liquid with a attribute wine-like odor and pungent style. In nature, grape-sugar breaks up by the motion of fermentation into alcohol or carbonic acid, with out something being added. As a psychoactive depressant, it's the lively ingredient in alcoholic drinks, and the second most consumed drug globally behind caffeine.
Ethanol is of course produced by the fermentation means of sugars by yeasts or by way of petrochemical processes equivalent to ethylene hydration. Traditionally it was used as a normal anesthetic, and has trendy medical purposes as an antiseptic, disinfectant, solvent for some medicines, and antidote for methanol poisoning and ethylene glycol poisoning. It's used as a chemical so
Web page: Alcohol by quantity
Abstract: Alcohol by quantity (abbreviated as alc/vol or ABV) is an ordinary measure of the quantity of alcohol contained in a given quantity of an alcoholic beverage, expressed as a quantity %. It's outlined because the variety of millilitres (mL) of pure ethanol current in 100 mL (3.5 imp fl oz; 3.4 US fl oz) of answer at 20 °C (68 °F). The variety of millilitres of pure ethanol is the mass of the ethanol divided by its density at 20 °C (68 °F), which is 0.78945 g/mL (0.82353 oz/US fl oz; 0.79122 oz/imp fl oz; 0.45633 oz/cu in). The alc/vol customary is used worldwide. The Worldwide Group of Authorized Metrology has tables of density of water–ethanol mixtures at completely different concentrations and temperatures.
In some nations, e.g. France, alcohol by quantity is also known as levels Homosexual-Lussac (after the French chemist Joseph Louis Homosexual-Lussac), though there's a slight distinction for the reason that Homosexual-Lussac conference makes use of the Worldwide Commonplace Environment worth for temperature, 15 °C (59 °F).
Web page: Alcohol gasoline
Abstract: Numerous alcohols are used as gasoline for inside combustion engines.  The primary 4 aliphatic alcohols (methanol, ethanol, propanol, and butanol)
are of curiosity as fuels as a result of they are often synthesized chemically or biologically, they usually have traits which permit them for use in inside combustion engines. The overall chemical system for alcohol gasoline is CnH2n+1OH.
Most methanol is produced from pure fuel, though it may be produced from biomass utilizing very related chemical processes. Ethanol is usually produced from organic materials by means of fermentation processes. Biobutanol has the benefit in combustion engines in that its vitality density is nearer to gasoline than the easier alcohols (whereas nonetheless retaining over 25% greater octane score); nevertheless, biobutanol is at present tougher to provide than ethanol or methanol. When obtained from organic supplies and/or organic processes, they're generally known as bioalcohols (e.g. "bioethanol"). There isn't any chemical dif
Web page: Cellulosic ethanol
Abstract: Cellulosic ethanol is ethanol (ethyl alcohol) produced from cellulose (the stringy fiber of a plant) quite than from the plant's seeds or fruit. It may be produced from grasses, wooden, algae, or different crops. It's typically mentioned to be used as a biofuel. The carbon dioxide that crops take in as they develop offsets among the carbon dioxide emitted when ethanol constituted of them is burned, so cellulosic ethanol gasoline has the potential to have a decrease carbon footprint than fossil fuels.
Curiosity in cellulosic ethanol is pushed by its potential to exchange ethanol constituted of corn or sugarcane. Since these crops are additionally used for meals merchandise, diverting them for ethanol manufacturing could cause meals costs to rise; cellulose-based sources, alternatively, typically don't compete with meals, for the reason that fibrous components of crops are principally inedible to people. One other potential benefit is the excessive variety and abundance of cellulose sources; grasses, bushes and algae are present in virtually each envi
AGENT:  Thought: The abstract of the web page 'Ethanol' mentions that ethanol is an alcohol with a attribute wine-like odor and pungent style. Nevertheless, it doesn't present the density of ethanol. I'll use the `search_info` device to seek out the density of ethanol from the 'Ethanol (knowledge web page)'.
Code:
```py
search_info('density of ethanol', 'Ethanol (knowledge web page)')
```
OBSERVATION:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Ethanol (knowledge web page)' for the question 'density of ethanol':
The density of ethanol at 20°C is roughly 0.7894 g/cm³.
AGENT:  Thought: I've discovered the density of ethanol from the Wikipedia web page 'Ethanol (knowledge web page)'. I'll return the end result utilizing the `final_answer` device.
Code:
```py
final_answer("The density of ethanol at 20°C is roughly 0.7894 g/cm³.")
```
OBSERVATION:  Print outputs:
Final output from code snippet:
The density of ethanol at 20°C is roughly 0.7894 g/cm³.

Web page search agent logs

TASK:  Retrieve details about the question:"density of birch plywood" from the Wikipedia web page "Plywood".
AGENT:  Thought: I would like to seek out the density of birch plywood. I'll use the device `retrieve_passages` to search for knowledge on the Wikipedia web page.
Code:
```py
end result = retrieve_passages("density of birch plywood")
print(end result)
```
OBSERVATION:  Print outputs:
Retrieved passages for question "density of birch plywood":
Passage 0: ... . Shear stress, in contrast, is greater within the middle of the panel, and on the outer fibres. Inside Europe primary plywood may be divided into three foremost classes: birch plywood (density approx. 680 kg/m3), combined plywood (density approx. 620 kg/m3) and conifer plywoods (density 460–520 kg/m3). ...
Passage 1: ... . Plane grade plywood is constituted of three or extra plies of birch, as skinny as 0.40 millimetres (1⁄64 in) thick in whole, and is extraordinarily sturdy and lightweight. ...
Passage 2: ... Baltic Birch plywood is a product of an space across the Baltic Sea. Initially manufactured for European cupboard makers however now well-liked in the USA as properly. It is extremely secure composed of an interior void-free core of cross-banded birch plys with an exterior grade adhesive. The face veneers are thicker than conventional cupboard grade ...
Passage 3: ... . Consequently, thicker panels can span higher distances underneath the identical masses. In bending, the utmost stress happens within the outermost layers, one in pressure, the opposite in compression. Bending stress decreases from the utmost on the face layers to almost zero on the central layer. Shear stress, in contrast, is greater within the middle of the panel, and on the outer fibres. Inside Europe primary plywood may be divided into three foremost classes: birch plywood (density approx ...
Passage 4: ... Wind turbine blades
Insulation packing containers for liquefied pure fuel (LNG) carriers
Clean floor and correct thickness mixed with the sturdiness of the fabric makes birch plywood a positive materials for a lot of particular finish makes use of e.g.:
Excessive-end loud audio system
Die-cutting boards
Supporting construction for parquet
Playground tools
Furnishings
Indicators and fences for demanding out of doors promoting
Musical devices
Sports activities tools...
AGENT:  Thought: From the passages, I can see that the density of birch plywood is roughly 680 kg/m³. I can use the `final_answer` device to supply the reply.
Code:
```py
final_answer("The density of birch plywood is roughly 680 kg/m³.")
```
OBSERVATION:  Print outputs:
Final output from code snippet:
The density of birch plywood is roughly 680 kg/m³.
------------------------------
TASK:  Retrieve details about the question:"density of ethanol" from the Wikipedia web page "Ethanol (knowledge web page)".
AGENT:  Thought: I would like to seek out the density of ethanol. I'll use the device `retrieve_passages` to get the data from the Wikipedia web page.
Code:
```py
end result = retrieve_passages("density of ethanol")
print(end result)
```
OBSERVATION:  Print outputs:
Retrieved passages for question "density of ethanol":
Passage 0: ... == Density of ethanol at numerous temperatures ==
Information obtained from Lange 1967
These knowledge correlate as ρ [g/cm3] = −8.461834×10−4 T [°C] + 0.8063372 with an R2 = 0.99999.
== Properties of aqueous ethanol options ==
Information obtained from Lange 1967
== Boiling factors of aqueous options ==
Information obtained from CRC Handbook of Chemistry (Web page 2117): 2391 
‡Azeotropic combination
== Charts ==
== References == ...
Passage 1: ... This web page gives supplementary chemical knowledge on ethanol.
== Materials Security Information Sheet ==
Exterior MSDS
== Construction and properties ==
== Thermodynamic properties ==
== Spectral knowledge ==
== Vapor strain of liquid ==
== Density of ethanol at numerous temperatures ==
Information obtained from Lange 1967
These knowledge correlate as ρ [g/cm3] = −8.461834×10−4 T [°C] + 0.8063372 with an R2 = 0.99999.
== Properties of aqueous ethanol options ==
Information obtained from Lange 1967...
AGENT:  Thought: From the passages, I can see that the density of ethanol varies with temperature. I can use the supplied linear equation to calculate the density at a particular temperature. I'll use the `math` module to carry out the calculation.
Code:
```py
# Extract the coefficients from the equation
a = -8.461834e-4
b = 0.8063372
# Calculate the density at a particular temperature, let's take 20°C for example
temperature = 20
density = a * temperature + b
print(density)
```
OBSERVATION:  Print outputs:
0.789413532
AGENT:  Thought: The passage gives a linear equation to calculate the density of ethanol at completely different temperatures. I can now use this equation to supply the density of ethanol at a particular temperature. I'll use the `final_answer` device to return the related info.
Code:
```py
final_answer(f"The density of ethanol at 20°C is roughly {spherical(0.789413532, 4)} g/cm³.")
```
OBSERVATION:  Print outputs:
Final output from code snippet:
The density of ethanol at 20°C is roughly 0.7894 g/cm³.

The mannequin appropriately identifies density distinction as the reason for floating/sinking, breaks up the duty into discovering the density of each substances, and is ready to attract the right conclusion from the retrieved knowledge regardless of them being in numerous items of measurement.

The mentioned system has spectacular capabilities, particularly contemplating the restricted measurement of the engine mannequin used, however after all, it has its personal limitations. The larger one appears to be associated to the facility of the engine mannequin itself: whereas it appears capable of perceive and observe complicated and lengthy directions in addition to to make use of the supplied instruments appropriately, typically it fails to adjust to easy tips and it tends to repeat unfruitful actions with out adjustments. That is significantly distinguished when the data the agent is on the lookout for isn’t current (or not simply extractible utilizing the semantic search device) from the supplied pages.

One other disadvantage, intrinsic to agentic and much more distinguished in multi-agentic programs, is the computation time wanted for the assorted brokers’ execution. In some easy instances the place the usual RAG workflow would have been sufficient to reply the query, this might produce an pointless computational overhead.

Lastly, having a number of brokers interacting with one another can amplify the danger of hallucinations, as a hallucination made by one agent might propagate over the execution chain, making it additionally very tough to trace.

Let’s visualize an instance by which the system, whereas nonetheless offering the right reply, made extreme use of device calls (with some repeated actions) unnecessarily extending the execution time.

Job: Which dynasty was ruling China on the time of the autumn of Constantinople?

The duty requires a very good degree of abstraction for the reason that two occasions usually are not associated and the precise 12 months of the autumn of Constantinople might not be reported on the web page of the corresponding dynasty.

The supervisor agent execution continues to be lean and on level, fixing the duty appropriately by first retrieving the date of the autumn of Constantinople after which on the lookout for Chinese language dynasties in that 12 months. Nevertheless, we are able to already see that the search outcomes, whereas nonetheless containing the required info, appear a bit unusual.

Agent:  Thought: To reply this query, I would like to seek out out when Constantinople fell and which Chinese language dynasty was ruling at the moment. I will begin by discovering the autumn date of Constantinople.
Code:
```py
wikipedia_search_agent(question="Fall of Constantinople")
```
Remark:  Print outputs:
Final output from code snippet:
The chief of the Ottoman Empire in 1453, throughout the fall of Constantinople, was Mehmed II, also referred to as Mehmed the Conqueror.
Agent:  Thought: From the earlier search, we all know that Constantinople fell in 1453, and it was conquered by Mehmed II. Now I would like to seek out out which Chinese language dynasty was ruling in 1453. I'll seek for the Chinese language dynasties round that point.
Code:
```py
wikipedia_search_agent(question="Chinese language dynasties 1453")
```
Remark:  Print outputs:
Final output from code snippet:
In 1453, the Ming dynasty was in full management of China. No particular occasions associated to 1453 are talked about within the info supplied.
Agent:  Thought: Primarily based on the data supplied, the Ming dynasty was in full management of China in 1453. Due to this fact, the dynasty ruling China when Constantinople fell could be the Ming dynasty.
Code:
```py
final_answer("The dynasty ruling China when Constantinople fell in 1453 was the Ming dynasty.")
```
Remark:  Print outputs:
Final output from code snippet:
The dynasty ruling China when Constantinople fell in 1453 was the Ming dynasty.

Let’s thus examine the corresponding Wikipedia search agent duties and their execution. We are able to see that, probably as a result of broad process “Fall of Constantinople”, the agent appears at different particular info after having discovered the right date. The identical occurs when retrieving details about the Chinese language dynasties, with the additional downside that some device calls have been repeated identically.

TASK:  Fall of Constantinople
AGENT:  Thought: I'll use the `search_wikipedia` device to seek out details about the autumn of Constantinople. It will seemingly present me with a abstract of the occasion and probably hyperlinks to associated pages.
Code:
```py
search_wikipedia("Fall of Constantinople")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Pages discovered for question 'Fall of Constantinople':
Web page: Fall of Constantinople
Abstract: The autumn of Constantinople, also referred to as the conquest of Constantinople, was the seize of the capital of the Byzantine Empire by the Ottoman Empire. The town was captured on 29 Might 1453 as a part of the end result of a 55-day siege which had begun on 6 April.
The attacking Ottoman Military, which considerably outnumbered Constantinople's defenders, was commanded by the 21-year-old Sultan Mehmed II (later nicknamed "the Conqueror"), whereas the Byzantine military was led by Emperor Constantine XI Palaiologos. After conquering the town, Mehmed II made Constantinople the brand new Ottoman capital, changing Adrianople.
The autumn of Constantinople and of the Byzantine Empire was a watershed of the Late Center Ages, marking the efficient finish of the Roman Empire, a state which started in roughly 27 BC and had lasted almost 1500 years. For a lot of trendy historians, the autumn of Constantinople marks the tip of the medieval interval and the start of the early trendy interval. The town's fall additionally stood as a turni
Web page: Sack of Constantinople
Abstract: The sack of Constantinople occurred in April 1204 and marked the end result of the Fourth Campaign. Crusaders sacked and destroyed most of Constantinople, the capital of the Byzantine Empire. After the seize of the town, the Latin Empire (identified to the Byzantines because the Frankokratia, or the Latin occupation) was established and Baldwin of Flanders topped as Emperor Baldwin I of Constantinople in Hagia Sophia.
After the town's sacking, many of the Byzantine Empire's territories have been divided up among the many Crusaders. Byzantine aristocrats additionally established a lot of small unbiased splinter states—one among them being the Empire of Nicaea, which might ultimately recapture Constantinople in 1261 and proclaim the reinstatement of the Empire. Nevertheless, the restored Empire by no means managed to reclaim all its former territory or attain its earlier financial power, and it regularly succumbed to the rising Ottoman Empire over the next two centuries.
The Byzantine Empire was left poorer, smal
Web page: Constantinople
Abstract: Constantinople (see different names) turned the capital of the Roman Empire throughout the reign of Constantine the Nice in 330. Following the collapse of the Western Roman Empire within the late fifth century, Constantinople remained the capital of the Jap Roman Empire (also referred to as the Byzantine Empire; 330–1204 and 1261–1453), the Latin Empire (1204–1261), and the Ottoman Empire (1453–1922). Following the Turkish Warfare of Independence, the Turkish capital then moved to Ankara. Formally renamed Istanbul in 1930, the town is right now the biggest metropolis in Europe, straddling the Bosporus strait and mendacity in each Europe and Asia, and the monetary middle of Turkey.
In 324, following the reunification of the Jap and Western Roman Empires, the traditional metropolis of Byzantium was chosen to function the brand new capital of the Roman Empire, and the town was renamed Nova Roma, or 'New Rome', by Emperor Constantine the Nice. On 11 Might 330, it was renamed Constantinople and devoted to Constantine. Constantin
Web page: Moscow, third Rome
Abstract: Moscow, third Rome (Russian: Москва — третий Рим; Moskva, tretiĭ Rim) is a theological and political idea asserting Moscow because the successor to historical Rome, with the Russian world carrying ahead the legacy of the Roman Empire. The time period "third Rome" refers to a historic subject of debate in European tradition: the query of the successor metropolis to the "first Rome" (Rome, inside the Western Roman Empire) and the "second Rome" (Constantinople, inside the Jap Roman Empire).
Web page: Byzantine Empire
Abstract: The Byzantine Empire, additionally known as the Jap Roman Empire, was the continuation of the Roman Empire centred in Constantinople throughout late antiquity and the Center Ages. The jap half of the Empire survived the situations that prompted the autumn of the West within the fifth century AD, and continued to exist till the autumn of Constantinople to the Ottoman Empire in 1453. Throughout most of its existence, the empire remained essentially the most highly effective financial, cultural, and navy drive within the Mediterranean world. The time period "Byzantine Empire" was solely coined following the empire's demise; its residents referred to the polity because the "Roman Empire" and to themselves as "Romans". Because of the imperial seat's transfer from Rome to Byzantium, the adoption of state Christianity, and the predominance of Greek as an alternative of Latin, trendy historians proceed to make a distinction between the sooner Roman Empire and the later Byzantine Empire.
Through the earlier Pax Romana interval, the western components of the empire b
AGENT:  Thought: From the abstract of the web page "Fall of Constantinople", I can see that the occasion occurred in 1453. I'll use the `search_info` device to seek out extra particulars concerning the occasion, particularly who was the chief of the Ottoman Empire at the moment.
Code:
```py
search_info("chief of the Ottoman Empire in 1453", "Fall of Constantinople")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Fall of Constantinople' for the question 'chief of the Ottoman Empire in 1453':
The chief of the Ottoman Empire in 1453 was Mehmed II, also referred to as Mehmed the Conqueror.
AGENT:  Thought: I've discovered that the chief of the Ottoman Empire in 1453 was Mehmed II, also referred to as Mehmed the Conqueror. I'll use the `final_answer` device to supply the reply.
Code:
```py
final_answer("The chief of the Ottoman Empire in 1453, throughout the fall of Constantinople, was Mehmed II, also referred to as Mehmed the Conqueror.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
The chief of the Ottoman Empire in 1453, throughout the fall of Constantinople, was Mehmed II, also referred to as Mehmed the Conqueror.
------------------------------
TASK:  Chinese language dynasties 1453
AGENT:  Thought: I'll use the `search_wikipedia` device to seek out details about Chinese language dynasties across the 12 months 1453. I'll present a normal question associated to Chinese language dynasties in that point interval.
Code:
```py
search_wikipedia('Chinese language dynasties 1453')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Pages discovered for question 'Chinese language dynasties 1453':
Web page: Chinese language Empire
Abstract: Chinese language Empire (or Empire of China) is a time period referring to the realm dominated by the Emperor of China throughout the period generally known as Imperial China. It was coined by western students and used to explain the Ming and Qing dynasties (or imperial Chinese language dynasties generally). One other time period was "Celestial Empire", in reference to the standing of the emperor because the Son of Heaven. In 221 BC, China was unified underneath an emperor for the primary time, and numerous dynasties or empires based by hereditary monarchs dominated China for a complete of two millennia since then, together with the Qin, Han, Jin, Sui, Tang, Tune, Yuan, Ming, and Qing.
Web page: Ming dynasty
Abstract: The Ming dynasty, formally the Nice Ming, was an imperial dynasty of China, ruling from 1368 to 1644 following the collapse of the Mongol-led Yuan dynasty. The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China. Though the first capital of Beijing fell in 1644 to a rebel led by Li Zicheng (who established the short-lived Shun dynasty), quite a few rump regimes dominated by remnants of the Ming imperial household—collectively referred to as the Southern Ming—survived till 1662.
The Ming dynasty's founder, the Hongwu Emperor (r. 1368–1398), tried to create a society of self-sufficient rural communities ordered in a inflexible, motionless system that might assure and help a everlasting class of troopers for his dynasty: the empire's standing military exceeded a million troops and the navy's dockyards in Nanjing have been the biggest on the earth. He additionally took nice care breaking the facility of the courtroom eunuchs and unrelated magnates, enfeoff
Web page: Listing of time durations
Abstract: The categorisation of the previous into discrete, quantified named blocks of time is named periodization. This can be a record of such named time durations as outlined in numerous fields of research.
These may be divided broadly into prehistorical durations and historic durations
(when written data started to be saved). 
In archaeology and anthropology, prehistory is subdivided into the three-age system, this record consists of using the three-age system in addition to a lot of numerous designation utilized in reference to sub-ages inside the conventional three.
The dates for every age can range by area. On the geologic time scale, the Holocene epoch begins on the finish of the final glacial interval of the present ice age (c. 10,000 BC) and continues to the current. The start of the Mesolithic is normally thought of to correspond to the start of the Holocene epoch.
Web page: Celestial Empire
Abstract: Celestial Empire (Chinese language: 天朝; pinyin: Tiāncháo; lit. 'heavenly dynasty') is an outdated title used to discuss with China or the Chinese language Empire, from a literary and poetic translation of the Chinese language time period, one among many names for China. The title was utilized in reference to the standing of the Emperor of China because the Son of Heaven within the Sinosphere.
Accordingly, within the nineteenth century, the title "Celestial" was used to discuss with Chinese language folks. Each phrases have been extensively used within the English-language well-liked mass media of the day, however fell into disuse in a while. Its utilization has develop into well-liked once more within the current day (2015), significantly amongst Chinese language Web customers. It's used to discuss with the present Communist regime, to indicate both disapproval for its political suppression and conceitedness or nationwide delight within the nation's emergence as a superpower within the twenty first century, relying on the context.
In trendy occasions it has principally fallen into disuse, however the title "Celestial Empire" nonetheless seems in numerous media equivalent to ar
Web page: Listing of dynasties
Abstract: This can be a record of monarchical dynasties and dynastic regimes organized by geographic area. Extant dynasties are rendered in daring and highlighted.
AGENT:  Thought: From the summaries, I can see that the Ming dynasty dominated from 1368 to 1644, which covers the 12 months 1453. I'll use the `search_info` device to seek out extra detailed details about the Ming dynasty throughout the 12 months 1453.
Code:
```py
search_info('1453', 'Ming dynasty')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Ming dynasty' for the question '1453':
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
AGENT:  Thought: Now that I do know the Ming dynasty dominated throughout 1453, I can present a extra particular reply concerning the occasions that occurred in China throughout that 12 months. Nevertheless, the data supplied doesn't give us particular particulars concerning the occasions in 1453. I'll use the `search_info` device to seek out extra detailed details about the Ming dynasty throughout the 12 months 1453.
Code:
```py
search_info('1453', 'Ming dynasty')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Ming dynasty' for the question '1453':
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
AGENT:  Thought: The knowledge supplied doesn't particularly point out the occasions of 1453. Nevertheless, we are able to infer that throughout the Ming dynasty, vital developments occurred. I'll use the `search_info` device to seek out extra particular details about the occasions of 1453 within the Ming dynasty.
Code:
```py
search_info('1453', 'Ming dynasty')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Ming dynasty' for the question '1453':
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
AGENT:  Thought: The knowledge supplied signifies that the Ming dynasty was in full management of China throughout the 12 months 1453. Nevertheless, there isn't any particular point out of occasions that occurred in 1453. I'll use the `search_info` device to seek out extra particular details about the occasions of 1453 within the Ming dynasty.
Code:
```py
search_info('1453', 'Ming dynasty')
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
Info retrieved from the web page 'Ming dynasty' for the question '1453':
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
AGENT:  Thought: The knowledge supplied doesn't particularly point out any occasions that occurred in 1453. Nevertheless, we all know that the Ming dynasty was in full management of China throughout this time. I'll use the `final_answer` device to supply the reply based mostly on the data we now have.
Code:
```py
final_answer("In 1453, the Ming dynasty was in full management of China. No particular occasions associated to 1453 are talked about within the info supplied.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
In 1453, the Ming dynasty was in full management of China. No particular occasions associated to 1453 are talked about within the info supplied.

Diving deeper into the web page search agent execution, we see that this additionally repeated a number of device calls with the identical parameters.

TASK:  Retrieve details about the question:"chief of the Ottoman Empire in 1453" from the Wikipedia web page "Fall of Constantinople".
AGENT:  Thought: I would like to seek out out who was the chief of the Ottoman Empire in 1453. I'll use the `retrieve_passages` device to seek for this info on the Wikipedia web page "Fall of Constantinople".
Code:
```py
end result = retrieve_passages("chief of the Ottoman Empire in 1453")
print(end result)
```
OBSERVATIONS:  Print outputs:
Retrieved passages for question "chief of the Ottoman Empire in 1453":
Passage 0: ... When Mehmed II succeeded his father in 1451, he was 19 years outdated. Many European courts assumed that the younger Ottoman ruler wouldn't severely problem Christian hegemony within the Balkans and the Aegean. In actual fact, Europe celebrated Mehmed coming to the throne and hoped his inexperience would lead the Ottomans astray. This calculation was boosted by Mehmed's pleasant overtures to the European envoys at his new courtroom. However Mehmed's gentle phrases weren't matched by his actions ...
Passage 1: ... The attacking Ottoman Military, which considerably outnumbered Constantinople's defenders, was commanded by the 21-year-old Sultan Mehmed II (later nicknamed "the Conqueror"), whereas the Byzantine military was led by Emperor Constantine XI Palaiologos. After conquering the town, Mehmed II made Constantinople the brand new Ottoman capital, changing Adrianople. ...
Passage 2: ... . The oldest boy, renamed Murad, turned a private favorite of Mehmed and served as Beylerbey (Governor-Basic) of Rumeli (the Balkans). The youthful son, renamed Mesih Pasha, turned Admiral of the Ottoman fleet and Sancak Beg (Governor) of the province of Gallipoli. He ultimately served twice as Grand Vizier underneath Mehmed's son, Bayezid II. ...
Passage 3: ... == Exterior hyperlinks ==
Media associated to Fall of Constantinople (1453) at Wikimedia Commons
The Siege of Constantinople As The Islamic World Sees it Archived 22 April 2021 on the Wayback Machine
World Historical past Encyclopedia – 1453: The Fall of Constantinople
Constantinople Siege & Fall, BBC Radio 4 dialogue with Roger Crowley, Judith Herrin & Colin Imber (In Our Time, 28 December 2006). ...
Passage 4: ... In the summertime of 1452, when Rumeli Hisarı was accomplished and the specter of the Ottomans had develop into imminent, Constantine wrote to the Pope, promising to implement the union, which was declared legitimate by a half-hearted imperial courtroom on 12 December 1452. Though he was longing for a bonus, Pope Nicholas V didn't have the affect the Byzantines thought he had over the Western kings and princes, a few of whom have been cautious of accelerating papal management...
AGENT:  Thought: From the passages, evidently the chief of the Ottoman Empire in 1453 was Mehmed II, also referred to as Mehmed the Conqueror. I'll use the `final_answer` device to supply the reply.
Code:
```py
final_answer("The chief of the Ottoman Empire in 1453 was Mehmed II, also referred to as Mehmed the Conqueror.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
The chief of the Ottoman Empire in 1453 was Mehmed II, also referred to as Mehmed the Conqueror.
------------------------------
TASK:  Retrieve details about the question:"1453" from the Wikipedia web page "Ming dynasty".
AGENT:  Thought: I would like to seek out info associated to the 12 months 1453 from the Wikipedia web page "Ming dynasty". I'll use the `retrieve_passages` device to get related passages.
Code:
```py
end result = retrieve_passages("1453 Ming dynasty")
print(end result)
```
OBSERVATIONS:  Print outputs:
Retrieved passages for question "1453 Ming dynasty":
Passage 0: ... The Ming dynasty, formally the Nice Ming, was an imperial dynasty of China, ruling from 1368 to 1644 following the collapse of the Mongol-led Yuan dynasty. The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China ...
Passage 1: ... === Reference works and first sources ===
Farmer, Edward L. ed. Ming Historical past: An Introductory Information to Analysis (1994).
Goodrich, Luther Carrington (1976). Dictionary of Ming Biography, 1368–1644. New York: Columbia College Press. ISBN 978-0-231-03833-1.
The Ming Historical past English Translation Undertaking, A collaborative mission that makes obtainable translations (from Chinese language to English) of parts of the 明史 Mingshi (Official Historical past of the Ming Dynasty). ...
Passage 2: ... ==== Reign of the Hongwu Emperor ==== ...
Passage 3: ... . The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China. Though the first capital of Beijing fell in 1644 to a rebel led by Li Zicheng (who established the short-lived Shun dynasty), quite a few rump regimes dominated by remnants of the Ming imperial household—collectively referred to as the Southern Ming—survived till 1662. ...
Passage 4: ... . Zhu Di assumed the throne because the Yongle Emperor (r. 1402–24); his reign is universally considered by students as a "second founding" of the Ming dynasty since he reversed a lot of his father's insurance policies....
AGENT:  Thought: The passages don't straight point out the 12 months 1453. Nevertheless, I can infer that the Ming dynasty lasted from 1368 to 1644, and 1453 falls inside this era. I'll use the `final_answer` device to supply a abstract based mostly on the data obtainable.
Code:
```py
final_answer("The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
------------------------------
TASK:  Retrieve details about the question:"1453" from the Wikipedia web page "Ming dynasty".
AGENT:  Thought: I would like to seek out info associated to the 12 months 1453 from the Wikipedia web page "Ming dynasty". I'll use the `retrieve_passages` device to get related passages.
Code:
```py
end result = retrieve_passages("1453 Ming dynasty")
print(end result)
```
OBSERVATIONS:  Print outputs:
Retrieved passages for question "1453 Ming dynasty":
Passage 0: ... The Ming dynasty, formally the Nice Ming, was an imperial dynasty of China, ruling from 1368 to 1644 following the collapse of the Mongol-led Yuan dynasty. The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China ...
Passage 1: ... === Reference works and first sources ===
Farmer, Edward L. ed. Ming Historical past: An Introductory Information to Analysis (1994).
Goodrich, Luther Carrington (1976). Dictionary of Ming Biography, 1368–1644. New York: Columbia College Press. ISBN 978-0-231-03833-1.
The Ming Historical past English Translation Undertaking, A collaborative mission that makes obtainable translations (from Chinese language to English) of parts of the 明史 Mingshi (Official Historical past of the Ming Dynasty). ...
Passage 2: ... ==== Reign of the Hongwu Emperor ==== ...
Passage 3: ... . The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China. Though the first capital of Beijing fell in 1644 to a rebel led by Li Zicheng (who established the short-lived Shun dynasty), quite a few rump regimes dominated by remnants of the Ming imperial household—collectively referred to as the Southern Ming—survived till 1662. ...
Passage 4: ... . Zhu Di assumed the throne because the Yongle Emperor (r. 1402–24); his reign is universally considered by students as a "second founding" of the Ming dynasty since he reversed a lot of his father's insurance policies....
AGENT:  Thought: The passages don't straight point out the 12 months 1453. Nevertheless, I can infer that the Ming dynasty lasted from 1368 to 1644, and 1453 falls inside this era. I'll use the `final_answer` device to supply a abstract based mostly on the data obtainable.
Code:
```py
final_answer("The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
------------------------------
TASK:  Retrieve details about the question:"1453" from the Wikipedia web page "Ming dynasty".
AGENT:  Thought: I would like to seek out info associated to the 12 months 1453 from the Wikipedia web page "Ming dynasty". I'll use the `retrieve_passages` device to get related passages.
Code:
```py
end result = retrieve_passages("1453 Ming dynasty")
print(end result)
```
OBSERVATIONS:  Print outputs:
Retrieved passages for question "1453 Ming dynasty":
Passage 0: ... The Ming dynasty, formally the Nice Ming, was an imperial dynasty of China, ruling from 1368 to 1644 following the collapse of the Mongol-led Yuan dynasty. The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China ...
Passage 1: ... === Reference works and first sources ===
Farmer, Edward L. ed. Ming Historical past: An Introductory Information to Analysis (1994).
Goodrich, Luther Carrington (1976). Dictionary of Ming Biography, 1368–1644. New York: Columbia College Press. ISBN 978-0-231-03833-1.
The Ming Historical past English Translation Undertaking, A collaborative mission that makes obtainable translations (from Chinese language to English) of parts of the 明史 Mingshi (Official Historical past of the Ming Dynasty). ...
Passage 2: ... ==== Reign of the Hongwu Emperor ==== ...
Passage 3: ... . The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China. Though the first capital of Beijing fell in 1644 to a rebel led by Li Zicheng (who established the short-lived Shun dynasty), quite a few rump regimes dominated by remnants of the Ming imperial household—collectively referred to as the Southern Ming—survived till 1662. ...
Passage 4: ... . Zhu Di assumed the throne because the Yongle Emperor (r. 1402–24); his reign is universally considered by students as a "second founding" of the Ming dynasty since he reversed a lot of his father's insurance policies....
AGENT:  Thought: The passages don't straight point out the 12 months 1453. Nevertheless, I can infer that the Ming dynasty lasted from 1368 to 1644, and 1453 falls inside this era. I'll use the `final_answer` device to supply a abstract based mostly on the data obtainable.
Code:
```py
final_answer("The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.
------------------------------
TASK:  Retrieve details about the question:"1453" from the Wikipedia web page "Ming dynasty".
AGENT:  Thought: I would like to seek out info associated to the 12 months 1453 from the Wikipedia web page "Ming dynasty". I'll use the `retrieve_passages` device to get related passages.
Code:
```py
end result = retrieve_passages("1453 Ming dynasty")
print(end result)
```
OBSERVATIONS:  Print outputs:
Retrieved passages for question "1453 Ming dynasty":
Passage 0: ... The Ming dynasty, formally the Nice Ming, was an imperial dynasty of China, ruling from 1368 to 1644 following the collapse of the Mongol-led Yuan dynasty. The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China ...
Passage 1: ... === Reference works and first sources ===
Farmer, Edward L. ed. Ming Historical past: An Introductory Information to Analysis (1994).
Goodrich, Luther Carrington (1976). Dictionary of Ming Biography, 1368–1644. New York: Columbia College Press. ISBN 978-0-231-03833-1.
The Ming Historical past English Translation Undertaking, A collaborative mission that makes obtainable translations (from Chinese language to English) of parts of the 明史 Mingshi (Official Historical past of the Ming Dynasty). ...
Passage 2: ... ==== Reign of the Hongwu Emperor ==== ...
Passage 3: ... . The Ming dynasty was the final imperial dynasty of China dominated by the Han folks, the bulk ethnic group in China. Though the first capital of Beijing fell in 1644 to a rebel led by Li Zicheng (who established the short-lived Shun dynasty), quite a few rump regimes dominated by remnants of the Ming imperial household—collectively referred to as the Southern Ming—survived till 1662. ...
Passage 4: ... . Zhu Di assumed the throne because the Yongle Emperor (r. 1402–24); his reign is universally considered by students as a "second founding" of the Ming dynasty since he reversed a lot of his father's insurance policies....
AGENT:  Thought: The passages don't straight point out the 12 months 1453. Nevertheless, I can infer that the Ming dynasty lasted from 1368 to 1644, and 1453 falls inside this era. I'll use the `final_answer` device to supply a abstract based mostly on the data obtainable.
Code:
```py
final_answer("The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.")
```
OBSERVATIONS:  Print outputs:
Final output from code snippet:
The 12 months 1453 falls inside the interval of the Ming dynasty, which dominated from 1368 to 1644. Throughout this time, the Ming dynasty was in full management of China and skilled vital developments equivalent to the development of the Forbidden Metropolis and the enlargement of the Nice Wall of China.

On this weblog publish, I defined the best way to create a multi-agentic RAG system utilizing code brokers and a “small” open-source LLM like Qwen2.5–7B-Instruct. I’ve mentioned the principle architectural options and a few particular decisions relative to the Hugging Face code agent implementation that I made to enhance the end result. The total code particulars can be found within the following GitHub repo.

The multi-agentic system described, regardless of being powered by a small mannequin operating on consumer-grade {hardware}, can remedy multi-hop question-answering duties associated to complicated queries. Specifically:

• It may possibly break down the question into manageable sub-tasks;
• It may possibly establish the Wikipedia pages containing the mandatory info;
• It may possibly mix info coming from a number of pages;
• It may possibly seek for detailed info on a Wikipedia web page;
• It may possibly decide whether or not it wants extra info and tries to seek out it;
• It may possibly efficiently repair small bugs within the code it produces and deal with device errors (like Wikipedia disambiguation errors).

I’ve additionally outlined some limitations of the system, equivalent to elevated computation time, repetitive actions, and the potential propagation of hallucinations. The latter may very well be mitigated by together with within the system a “proofreader” agent that checks that the reported info is in settlement with the retrieved sources.

Additionally it is value noting that, for the reason that agentic system has an ordinary RAG strategy at its core, all the standard strategies used to enhance the effectivity and accuracy of the latter may be applied within the framework.

One other attainable enchancment is to make use of strategies to extend take a look at time computation to present the mannequin extra “time to assume” much like OpenAI o1/o3 fashions. It’s nevertheless essential to notice that this modification will additional enhance execution time.

Lastly, for the reason that multi-agentic system is made up of brokers specialised in a single process, utilizing a unique mannequin engine for every of them might enhance the efficiency. Specifically, it’s attainable to fine-tune a unique mannequin for every process within the system for additional efficiency positive aspects. This may very well be significantly useful for small fashions. It’s value mentioning that fine-tuning knowledge may be collected by operating the system on a set of predetermined duties and saving the brokers’ output when the system produces the right reply, thus eliminating the necessity for costly handbook knowledge annotation.

I hope you discovered this tutorial helpful, yow will discover the total code implementation within the GitHub repo and take a look at it your self within the Colab pocket book.