Bigger Context Home windows Don’t Repair RAG — So I Constructed a System That Does

TL;DR

• I constructed a dataset Q&A system and trusted a RAG reply that was lower than half-correct.
• I measured this throughout 7 question sorts and 5 context sizes on 100,000 rows.
• queries away from RAG completely.

I Trusted the Incorrect Quantity

Final month I used to be heads-down constructing a brand new function for EmiTechLogic. Learners can now add their very own messy CSV information and ask questions in plain English about their information. Sounded good for RAG, so I went all in — embeddings, retrieval, nice-looking responses.

The primary few demos seemed superb. Clear tables, assured numbers, skilled formatting. I truly began trusting the system in our inside testing.

Then I picked one quantity to double-check.

Actual grocery spend within the dataset: $1,140,033.24.  

The mannequin gave me a fantastic breakdown by class. It seemed legit. I added up the numbers it returned.

It was lower than half.

I sat there staring on the display screen considering “this will’t be proper.” So I did what any engineer would do. I elevated the context window. 4k… 16k… 32k… 128k tokens. Every time the reply received longer, extra detailed, and extra confidently incorrect.

That’s when it lastly clicked. This wasn’t a retrieval subject. I used to be asking a retrieval system to carry out heavy computation on information it had solely partially seen. And as a substitute of claiming it was not sure or lacking info, the mannequin was producing polished, structured solutions that seemed right.

Why RAG Can’t Mixture

The RAG pipeline doesn’t really perceive structured information. All it does is take every CSV row and flatten it into plain textual content. That’s it. To the mannequin, a row appears to be like one thing like this:

"2019-01-01 grocery_pos 107.23 F NC Jennifer Banks ..."

For a question like “What’s the whole spend by class?”, the RAG pipeline does this:

1. Tokenise: ["total", "spend", "category"]
2. Rating all 100,000 rows by key phrase overlap
3. Return the top-N rows as serialised plain textual content
4. Ask the LLM to sum and group from that textual content

Step 4 is the place the system fails. The LLM shouldn’t be working a SUM. It’s pattern-matching numbers from a textual content blob and producing a response that mimics an aggregation.

Fashions battle with numerical precision at scale [1], however the true subject is the presentation. The mannequin offers you an in depth breakdown throughout all classes. It is a traditional entice. The output appears to be like skilled. It mimics the construction of an actual report so properly that your mind assumes the content material is legitimate. You haven’t any method to confirm that 92% of your information is lacking.

RAG is a retrieval instrument. It’s not a calculation engine. Retrieval finds related fragments. Computation requires a full dataset scan. While you use RAG for math, you get a incorrect reply that appears authoritative. That distinction is vital. A partial reply alerts that information is lacking. An entire-looking incorrect reply simply alerts false confidence.

Full code: https://github.com/Emmimal/context-window-engine/

The Benchmark: Two Pipelines, Similar Question

To measure this exactly, I constructed a benchmark that runs two pipelines facet by facet for each question.

The primary pipeline is a RAG simulation. It fashions what a naive vector pipeline passes to an LLM at 5 context sizes. I examined 5 context sizes, starting from 5 rows as much as 8,000. That scales from 325 tokens to 500,000. For every dimension, I tracked three metrics: how a lot information the LLM sees, what sum it computes from that particular slice, and whether or not a reader might truly spot the error.

The second pipeline is a semantic engine that executes the identical question as a deterministic full-scan over all 100,000 rows and returns the precise right reply.

The simulation doesn’t reproduce precise LLM outputs. What it preserves is the important thing structural property: a partial slice of information fed right into a system that returns a full-form reply. That’s the property that causes the issue, and that’s what the benchmark measures.

I selected seven question sorts to cowl each aggregation sample a structured information system is prone to encounter:

These queries aren’t distinctive or complicated. They’re the usual questions any analyst asks when taking a look at a brand new dataset. That’s precisely why this failure is so vital.

Error Observability Collapse

Right here is the total benchmark output for the question that began all of this. I’m exhibiting it in full as a result of the numbers make the issue unimaginable to dismiss.

GROUND TRUTH (Semantic Engine)
SUM(amt) GROUP BY class → 14 teams
  #1  grocery_pos               1,140,033.24
  #2  shopping_net                773,527.93
  #3  shopping_pos                725,766.14
  #4  gas_transport               648,804.24
  #5  residence                        556,526.53
Latency: 100.47ms | Rows scanned: 100,000

RAG SIMULATION — what the LLM receives at every context dimension

Context               Rows   Protection    Partial sum  Error detectable?
tiny   (~325 tokens)     5   0.0050%         197.73  EASY
small  (~3K tokens)     50   0.0500%       2,003.56  MODERATE
medium (~32K tokens)   500   0.5000%      31,023.21  HARD
massive  (~130K tokens) 2,000  2.0000%     140,093.16  VERY HARD
xlarge (~520K tokens) 8,000  8.0000%     569,368.22  NEAR IMPOSSIBLE

I stared at these outcomes for some time. Essentially the most troubling half wasn’t simply that the solutions have been incorrect, it was how a lot tougher the errors grew to become to identify because the context window grew.

At 8,000 rows the error was nonetheless over 50%, but the response seemed like an expert report. You’d must manually confirm the numbers to note one thing was off. That’s what I began calling Error Observability Collapse. The extra context I gave the mannequin, the extra convincing — however no more correct — the output grew to become.

The “Partial sum” column exhibits the overall if the LLM added each quantity worth within the rows it truly retrieved. The “Error detectable?” column scores how seemingly a human reader is to identify a mistake.

With 5 rows, the partial sum is 197.73. The proper whole is 1,140,033.24. It’s apparent. The output is brief, the numbers are incorrect, and the lacking information is evident. The error is immediate.

At 8,000 rows, the partial sum hits 569,368.22. The LLM has now seen all 14 classes. It generates a 1,500-word report with particular figures and assured language. The error is 50%, however it’s hidden inside authoritative, well-structured prose. With out an exterior reference, a reader has no method to catch it.

That is the sample that held throughout all seven queries:

I known as this Error Observability Collapse. As context grows, confidence scales with it. Correctness doesn’t.

The failure modes are uneven, which makes them harmful:

A incorrect RAG reply appears to be like right. It’s formatted, particular, and assured. A failed computation throws an express error. It’s seen.

One failure is silent. The opposite is loud. As context home windows attain tens of millions of tokens, the silent failure turns into tougher to detect [4]. The system doesn’t get safer because it scales. It simply will get extra convincing.

The Semantic Engine: Proof That the Appropriate Reply Is Quick

Earlier than I absolutely understood the issue, I had already thrown collectively a easy semantic engine out of frustration. I simply wished the right reply at the very least as soon as.

The method turned out to be easy: parse the question into correct operations and run a single cross over your entire dataset. No embeddings, no retrieval, no guessing.

Right here’s what that appears like in follow:

The logic is easy. Take a question like “What’s the whole spend by class?”. The engine maps this to a direct operation: SUM(amt) GROUP BY class. It processes the total 100,000-row set in a single cross. It accumulates grouped totals. There isn’t a retrieval. No inference. No partial scanning. It visits each row as soon as and returns the precise outcome.

This proves that the right reply shouldn’t be costly. Benchmark queries completed below 200ms. Pattern dimension: 100,000 rows. Aggregation is trivial. The failure occurs once you route these queries to a system constructed to misconceive them.

from context_window_engine import compute_ground_truth, load_csv

rows = load_csv("information/credit_card_transactions.csv", max_rows=100_000)

gt = compute_ground_truth(
    query_label = "whole by class",
    rows        = rows,
    agg_func    = "sum",
    agg_col     = "amt",
    group_col   = "class",
)
# gt.reply     → [(grocery_pos, 1140033.24), (shopping_net, 773527.93), ...]
# gt.latency_ms → 100.47

Engine helps SUM, AVG, COUNT, MIN, MAX. Handles categorical and numeric filters. Contains GROUP BY and ratios. Zero exterior dependencies. Each operation runs as a deterministic operate over the total listing.

The engine itself isn’t the product. It’s the proof: the right reply is reachable below a second. No inference required. The actual problem is routing queries there reliably.

The Repair Is Not Higher Retrieval

Cease attempting to enhance retrieval. If a question wants 100% of the info, an 8% pattern fails. The repair is eradicating retrieval from the loop.

We want a classification layer. It sits earlier than the pipeline and makes one binary name: computation or lookup?

The distinction is evident. “Complete spend by class” calls for a full scan. “Discover transactions from Jennifer Banks” is an easy lookup. Commonplace RAG forces each down the identical path. That’s the design flaw.

A QueryRouter fixes this. It inspects each incoming question and routes it to the right path earlier than a single retrieval begins.

The classifier makes use of three sign tiers, prioritized. Tier 1: aggregation verbs—whole, what number of, common, lowest, proportion. These demand full-dataset computation. Tier 2: numeric comparability—better than 500, above $1,000, at the very least. These suggest filter-then-aggregate, unimaginable for RAG. Tier 3: retrieval alerts—discover, present me, listing, fetch. These point out lookups the place semantic similarity works.

Default to COMPUTATION if no tier matches. That is deliberate. Failure modes are uneven: a incorrect RAG reply on an aggregation is silently incorrect. A computation engine that may’t parse a question throws an error. When unsure, fail loudly.

from query_router import QueryRouter

router = QueryRouter(rows)

outcome = router.route("What's the whole spend by class?")
# outcome.routed_to     → "COMPUTATION"
# outcome.reply.reply → [(grocery_pos, 1140033.24), ...]
# outcome.total_latency → ~250ms — classify + execute mixed

outcome = router.route("Discover transactions from Jennifer Banks")
# outcome.routed_to     → "RETRIEVAL"
# outcome.reply.secure   → True — RAG is acceptable

Routing the Full Benchmark

I ran 9 queries by means of the router to confirm efficiency throughout each sorts: seven aggregation queries destined for the semantic engine, and two lookup queries for RAG.

Each route was right. The seven aggregation queries hit the full-scan engine and returned precise outcomes. The 2 lookup queries appropriately triggered the RAG path. Have a look at the output: excessive confidence scores, right sample matching, and latency below 130ms—even with the 100,000-row scan.

[1] ✓  COMPUTATION   "What's the whole spend by class?"
     Tier 1 | matched='whole' | confidence=0.97
     #1 grocery_pos      1,140,033.24  (102.57ms | 100,000 rows | precise)

[2] ✓  COMPUTATION   "Which class has the best common transaction quantity?"
     Tier 1 | matched='highest' | confidence=0.97
                               71.91  (119.47ms | 100,000 rows | precise)

[3] ✓  COMPUTATION   "What's the whole quantity spent on grocery_pos?"
     Tier 1 | matched='whole' | confidence=0.97
                        1,140,033.24  (49.96ms  | 100,000 rows | precise)

[4] ✓  COMPUTATION   "What number of transactions have been made by feminine prospects?"
     Tier 1 | matched='What number of' | confidence=0.97
                           54,641.00  (90.45ms  | 100,000 rows | precise)

[5] ✓  COMPUTATION   "What's the whole spend the place quantity is bigger than 500?"
     Tier 1 | matched='whole' | confidence=0.97
                        1,274,269.60  (91.65ms  | 100,000 rows | precise)

[6] ✓  COMPUTATION   "Which state has the bottom whole spending?"
     Tier 1 | matched='lowest' | confidence=0.97
     lowest RI               2,125.60  (109.05ms | 100,000 rows | precise)

[7] ✓  COMPUTATION   "What proportion of transactions are fraudulent?"
     Tier 1 | matched='proportion' | confidence=0.97
                              0.9900%  (87.35ms  | 100,000 rows | precise)

[8] ✓  RETRIEVAL     "Discover transactions from Jennifer Banks"
     Tier 3 | matched='Discover' | confidence=0.85
     RAG is acceptable — no aggregation required

[9] ✓  RETRIEVAL     "Present me a pattern transaction from Texas"
     Tier 3 | matched='Present me' | confidence=0.85
     RAG is acceptable — no aggregation required

Routing accuracy: 9/9

9/9 right. Error Observability Collapse is unimaginable if aggregation queries by no means attain RAG.

The Check Suite

The benchmark verifies 9 particular queries. The take a look at suite ensures reliability throughout a broader vary: edge circumstances, malformed inputs, lacking information, and customary manufacturing failure factors.

The engine suite has 87 checks throughout 10 courses. It covers float parsing with greenback indicators, commas, and scientific notation; all 5 aggregation capabilities below regular circumstances and with empty inputs; all 5 numeric filter operators; full GROUP BY aggregation with categorical and numeric filters mixed; RAG simulation protection metrics at every context dimension; and edge circumstances together with empty datasets, rows with lacking column values, and single-row inputs.

The router suite has 72 checks throughout 5 courses. It covers all three tier patterns, together with edge circumstances like all-caps queries and really lengthy queries; pure language to typed operation parsing for each supported question kind; routing and execution correctness towards all seven benchmark queries; and a distinction suite that verifies router solutions match unbiased ground-truth computation — guaranteeing the router doesn’t introduce any deviation from the engine’s personal output.

Run the engine checks by typing python area -m area unittest area test_engine area -v. This executes the 87 checks within the suite.

Run the router checks by typing python area -m area unittest area test_router area -v. This executes the 72 checks within the suite.

All 159 cross on Python 3.9+ with zero exterior dependencies.

Trustworthy Limitations

This resolution isn’t good. It solely works on single CSV information proper now. Actual manufacturing datasets are normally messy with a number of tables that want becoming a member of — I intentionally stored the scope small as a result of I wished one thing that really labored end-to-end first.

The router can also be nonetheless fairly fundamental (regex-based). I attempted a small LLM-based classifier early on however it was inconsistent and added latency, so I went again to the easy method. Generally the boring resolution wins.

I additionally simulated the RAG responses as a substitute of hitting actual APIs for the benchmark. The patterns maintain up, however your mileage with GPT-4o or Claude 3.5 will fluctuate barely.

CSV format required. The engine hundreds structured information immediately from CSV information. Database connections, Parquet information, and different tabular codecs aren’t supported right now.

What This Modifications

Including a routing layer prices virtually nothing. Classifying a question towards 65 regex patterns takes simply microseconds. The semantic engine provides lower than 200ms to scan a 100,000 row dataset. The full overhead is smaller than a single embedding name.

What you get in return is a deterministic reply for each aggregation question. Each whole, each depend, and each proportion now comes from a full scan as a substitute of a assured approximation based mostly on 8 % of the info. RAG retains dealing with what it’s truly good at: retrieving particular information, surfacing related passages, and answering lookup questions the place semantic similarity is the appropriate instrument for the job.

RAG shouldn’t be damaged. It’s simply being requested to compute, and it can not try this.
The damaging half shouldn’t be that it fails. It’s that it fails convincingly. And no quantity of context adjustments that.

You possibly can strive typing it out like this:

To begin, clone the repository utilizing git clone adopted by the URL https://github.com/Emmimal/context-window-engine/. As soon as that finishes, transfer into the listing by typing cd context-window-engine. Lastly, launch the venture by working python demo.py in your terminal.

References

[1] Levy, M., Jacoby, A., & Goldberg, Y. (2024). Similar activity, extra tokens: The influence of enter size on the reasoning efficiency of enormous language fashions. In Proceedings of the 62nd Annual Assembly of the Affiliation for Computational Linguistics (Quantity 1: Lengthy Papers), pages 15339–15353, Bangkok, Thailand. Affiliation for Computational Linguistics.
https://doi.org/10.18653/v1/2024.acl-long.818

[2] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-t., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented era for knowledge-intensive NLP duties. Advances in Neural Data Processing Methods, 33, 9459–9474. https://doi.org/10.48550/arXiv.2005.11401

[3] Gao, Y., Xiong, Y., Gao, X., Jia, Ok., Pan, J., Bi, Y., Dai, Y., Solar, J.,
Guo, Q., Wang, M., & Wang, H. (2023). Retrieval-augmented era for big language fashions: A survey. arXiv preprint arXiv:2312.10997.
https://doi.org/10.48550/arXiv.2312.10997

[4] Liu, N. F., Lin, Ok., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2023). Misplaced within the center: How language fashions use lengthy contexts. Transactions of the Affiliation for Computational Linguistics, 12, 157–173. https://doi.org/10.1162/tacl_a_00638

[5] Koshorek, O., Granot, N., Alloni, A., Admati, S., Hendel, R., Weiss, I., Arazi, A., Cohen, S.-N., & Belinkov, Y. (2025). Structured RAG for answering aggregative questions. arXiv preprint arXiv:2511.08505.
https://doi.org/10.48550/arXiv.2511.08505

Disclosure

All benchmark numbers are from precise runs on Python 3.12.6, Home windows 11, CPU solely, no GPU. The benchmark makes use of the Credit score Card Transactions Fraud Detection dataset (Kartik Gajjar, Kaggle, 2020), an artificial dataset generated utilizing the Sparkov transaction simulator created by Brandon Harris, licensed CC0 (Public Area), accessible at kaggle.com/datasets/kartik2112/fraud-detection. The RAG baseline simulates retrieval and fashions confidence alerts — no actual LLM API calls are made. No exterior API keys are required to breed any outcome on this article. All code described right here was written and examined by me.