RAG Isn’t Enough — I Built the Missing Context Layer That Makes LLM Systems Work

TL;DR

a full working implementation in pure Python, with actual benchmark numbers.

RAG techniques break when context grows past a couple of turns.

The true downside is just not retrieval — it’s what really enters the context window.

A context engine controls reminiscence, compression, re-ranking, and token limits explicitly.

This isn’t an idea. It is a working system with measurable habits.

The Breaking Level of RAG Methods

I constructed a RAG system that labored completely — till it didn’t.

The second I added dialog historical past, all the things began breaking. Related paperwork have been getting dropped. The immediate overflowed. The mannequin began forgetting issues it had mentioned two turns in the past. Not as a result of retrieval failed. Not as a result of the immediate was badly written. However as a result of I had zero management over what really entered the context window.

That’s the issue no one talks about. Most RAG tutorials cease at: retrieve some paperwork, stuff them right into a immediate, name the mannequin. What occurs when your retrieved context is 6,000 characters however your remaining finances is 1,800? What occurs when three of your 5 retrieved paperwork are near-duplicates, crowding out the one helpful one? What occurs when flip one among a twenty-turn dialog remains to be sitting within the immediate, taking on area, lengthy after it stopped being related?

These aren’t uncommon edge circumstances. That is what occurs by default — and it begins breaking inside the first few turns.

All outcomes under are from actual runs of the system (Python 3.12, CPU-only, no GPU), besides the place famous as calculated.

The reply is a layer most tutorials skip completely. Between uncooked retrieval and immediate building, there’s a deliberate architectural step: deciding what the mannequin really sees, how a lot of it, and in what order. In 2025, Andrej Karpathy gave this a reputation: context engineering [2]. I’d been constructing it for months with out calling it that.

That is the system I constructed from retrieval to reminiscence to compression with actual numbers and code you’ll be able to run.

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

What Context Engineering Really Is

It’s value being exact, as a result of the phrases get muddled.

Immediate engineering is the craft of what you say to the mannequin — your system immediate, your few-shot examples, your output format directions. It shapes how the mannequin causes.

RAG is a method for fetching related exterior paperwork and together with them earlier than era. It grounds the mannequin in info it wasn’t educated on [1].

Context engineering is the layer in between — the architectural choices about what data flows into the context window, how a lot of it, and in what type. It solutions: given all the things that would go into this immediate, what ought to really go in?

All three are complementary. In a well-designed system they every have a definite job.

Who This Is For

This structure is value constructing in case you are engaged on multi-turn chatbots the place context accumulates throughout turns, RAG techniques with massive data bases the place retrieval noise is an actual downside, or AI copilots and brokers that want reminiscence to remain coherent.

Skip it for single-turn queries with a small data base — the pipeline overhead doesn’t justify a marginal high quality acquire. Skip it for latency-critical providers underneath 50ms — embedding era alone provides ~85ms on CPU. Skip it for totally deterministic domains like authorized contract evaluation, the place keyword-only retrieval is commonly ample and extra auditable.

When you’ve got limitless context home windows and limitless latency, plain RAG works positive. In manufacturing, these constraints don’t exist.

Full Pipeline Structure

Element 1: The Retriever

Most RAG implementations decide one retrieval methodology and name it performed. The issue isn’t any single methodology dominates throughout all question sorts. Key phrase matching is quick and exact for actual phrases. TF-IDF handles time period weighting. Dense vector embeddings catch semantic relationships that key phrases miss completely.

Key phrase vs. TF-IDF — Identical Question, Totally different Habits

For the question: “how does reminiscence work in AI brokers”

Each strategies agree on mem-001 as the highest doc. However there’s a important distinction: TF-IDF supplies extra nuanced scoring by weighting time period rarity, whereas key phrase retrieval solely counts uncooked overlap. On this question they converge — however they diverge badly on conceptual queries with totally different wording. That is exactly why hybrid retrieval turns into vital.

The Retriever helps three modes: key phrase, tfidf, and hybrid. Hybrid mode runs each strategies and blends their scores with a single tunable weight:

hybrid_score = alpha * emb_score + (1 - alpha) * tf_score

The alpha=0.65 default weights embeddings barely greater than TF-IDF — empirical, not principled, however examined throughout totally different question types. Key phrase-heavy queries carry out higher round alpha=0.4; paraphrase-style queries profit from alpha=0.8 or larger.

What Hybrid Retrieval Fixes That TF-IDF Misses

For the question: “how do embeddings evaluate to TF-IDF for reminiscence in AI brokers”

tfidf-001 doesn’t seem in TF-IDF outcomes as a result of it shares few question tokens. Hybrid mode surfaces it as a result of the embedding recognises its conceptual relevance. That is the precise failure mode of conventional RAG at scale.

One implementation notice: sentence-transformers is elective. With out it, the system falls again to random embeddings with a warning. Manufacturing will get actual semantics; improvement will get a purposeful stub.

Element 2: The Re-ranker

Retrieval provides you candidates. Re-ranking decides the ultimate order.

The re-ranker applies a two-factor weighted sum mixing retrieval rating with a tag-based significance worth. Paperwork tagged with reminiscence, context, rag, or embedding obtain a tag_importance of 1.4; all others obtain 1.0. Each feed into the identical formulation:

final_score = base_score * 0.68 + tag_importance * 0.32

A tagged doc with tag_importance=1.4 contributes 0.448 from that time period alone, versus 0.32 for an untagged one — a hard and fast bonus of 0.128 no matter retrieval rating. The weights replicate a selected prior: retrieval sign is major, area relevance is a significant secondary sign.

Scores Earlier than and After Re-ranking

rag-001 jumps from exterior the highest 4 to second place completely as a consequence of its tag enhance. These reorderings change which paperwork survive compression — they’re not beauty.

Is the heuristic principled? Not completely. A cross-encoder re-ranker — scoring every query-document pair with a neural mannequin [7] — can be extra correct. However cross-encoders value one mannequin name per doc. At 5 paperwork, the heuristic runs in microseconds. At 500+, a cross-encoder turns into value the price.

Element 3: Reminiscence with Exponential Decay

That is the part most tutorials miss completely, and the one the place naive techniques collapse quickest.

Conversational reminiscence has two failure modes: forgetting too quick (shedding context that’s nonetheless related) and forgetting too gradual (accumulating noise that crowds out helpful data). A sliding window drops outdated turns abruptly — flip 10 is totally current, flip 11 is gone. That’s not how helpful data works.

The answer is exponential decay, the place turns fade repeatedly based mostly on three elements.

The scoring formulation:

efficient = significance * recency * freshness + relevance_boost

The place every time period is:

• recency = e^(−decay_rate × age_seconds) — older turns carry much less weight
• freshness = e^(−0.01 × time_since_last_access) — just lately referenced turns get a lift
• relevance_boost = (|question ∩ flip| / |question|) × 0.35 — turns with excessive query-token overlap are retained longer

This mirrors how working reminiscence really prioritises data [4] — high-importance turns survive longer; off-topic turns fade rapidly no matter after they occurred.

Auto-Significance Scoring

Auto-importance scoring makes this sensible with out handbook annotation. The system scores every flip based mostly on content material size, area key phrases, and question overlap:

A climate query scores 1.10 — barely above the ground. A website query about reminiscence decay scores 2.50 and survives far longer earlier than decaying. In a protracted dialog, high-importance area turns keep in reminiscence whereas low-importance small-talk turns fade first — the precise ordering you need.

Deduplication

Deduplication runs earlier than any flip is saved, as a three-tier verify: actual containment (if the brand new flip is a substring of an current one, reject), robust prefix overlap (if the primary half of each turns match, reject), and token-overlap similarity >= 0.72 (if token overlap is excessive sufficient, reject as a paraphrase).

At 0.72, you catch paraphrases with out falsely rejecting related-but-distinct questions on the identical subject. A follow-up like “Are you able to clarify context engineering and its position in RAG?” after “What’s context engineering and the way does it assist RAG techniques?” scores ~72% overlap — deduplication fires, one reminiscence slot saved, room made for genuinely new data.

Token Funds Underneath Stress

Element 4: Context Compression

You’ve gotten 810 characters of retrieved context. Your remaining token finances permits 800. That 10-character hole means one thing both will get truncated badly or the entire thing overflows.

The Compressor implements three methods. Truncate is the quickest — cuts every chunk proportionally. Sentence makes use of grasping sentence-boundary choice. Extractive is query-aware: each sentence throughout all retrieved paperwork will get scored by token overlap with the question, ranked by relevance, and greedily chosen inside finances. Then the chosen sentences are served again of their authentic doc order, not relevance rank order [5]. Relevance rank order produces incoherent context. Authentic order preserves the logical movement of the supply materials.

Compression Technique Commerce-offs — Identical 810-Character Enter, 800-Character Funds

Extractive compression preserves which means higher — however saves fewer uncooked characters. Underneath tight budgets, it provides you the proper content material, not simply much less content material.

Element 5: The Token Funds Enforcer

Every little thing feeds into the TokenBudget — a slot-based allocator that tracks utilization throughout named context areas. Token estimation makes use of the 1 token ≈ 4 characters heuristic for English prose, in keeping with OpenAI’s documentation [6].

The order of reservation is the entire design:

def construct(self, question: str) -> ContextPacket:
    finances = TokenBudget(complete=self.total_token_budget)
    finances.reserve_text("system_prompt", self.system_prompt)          # 1. Fastened

    scored_docs = self._rerank(self._retriever.retrieve(question, ...), question)

    memory_turns = self._memory.get_weighted(question=question)
    finances.reserve_text("historical past", " ".be a part of(t.content material for t in memory_turns))  # 2. Reserved

    remaining_chars = finances.remaining_chars()
    compressor = Compressor(max_chars=remaining_chars, technique=self.compression_strategy)
    consequence = compressor.compress([sd.document.content for sd in scored_docs], question=question)

    finances.reserve_text("retrieved_docs", consequence.textual content)                # 3. What's left
    return ContextPacket(...)

The system immediate is fastened overhead you’ll be able to’t negotiate away. Reminiscence is what makes multi-turn coherent. Paperwork are the variable — helpful, however the very first thing to compress when area runs out. Reserve within the incorrect order and paperwork silently overflow the finances earlier than historical past is even accounted for. The orchestrator enforces the proper order explicitly.

What Occurs Underneath Actual Token Stress

That is the place naive techniques fail — and this engine adapts.

Setup: 5 paperwork (810 chars complete), 200 tokens reserved for system immediate, 800-token complete finances. Question: “How do embeddings and TF-IDF evaluate for reminiscence in brokers?”

Flip 1 — no dialog historical past but: Paperwork retrieved: 5, re-ranked. Reminiscence turns: 0. Compression utilized: 48% discount. Outcome: matches inside finances.

Flip 2 — after dialog begins: Paperwork retrieved: 5, re-ranked. Reminiscence turns: 2, now competing for area. Compression turns into extra aggressive: 45% discount. Outcome: nonetheless matches inside finances.

What modified? The system didn’t fail — it tailored. Reminiscence turns consumed a part of the finances, so compression on retrieved paperwork tightened robotically. That’s the purpose of context engineering: the mannequin at all times receives one thing coherent, by no means a random overflow.

Measuring What It Really Buys You

The desk under compares 4 approaches on the identical question and 800-token complete finances. The primary three rows are calculated from identified inputs utilizing the identical 810-character doc set; the fourth row displays precise engine output verified towards demo runs.

Naive RAG overflows instantly. Truncation matches however blindly cuts the tail. Reminiscence with out decay provides noise relatively than sign — older turns by no means fade, and dialog historical past turns into bloat. The complete system re-ranks, compresses intelligently, and consists of solely turns that also carry data.

Reminiscence Decay by Significance Rating

Efficiency Traits

Measured on Python 3.12.6, CPU solely, no GPU, 5-document data base:

Hybrid retrieval is the bottleneck. Should you want sub-50ms response time, use TF-IDF or key phrase mode as an alternative. At 100 requests/sec in hybrid mode you want roughly 9 concurrent staff; with embedding caching, subsequent queries drop to ~2ms per request after the primary.

Sincere Design Selections

alpha=0.65 is empirical, not principled. I examined throughout a small question set from my data base. For a special area — authorized paperwork, medical literature, dense code — the proper alpha will likely be totally different. Key phrase-heavy queries do higher round 0.4; conceptual or paraphrased queries profit from 0.8 or larger.

The re-ranking weights (0.68/0.32) are a heuristic. A cross-encoder re-ranker can be extra principled [7] however prices one mannequin name per doc. For five paperwork, the heuristic runs in microseconds. For 500+ paperwork, a cross-encoder turns into value the price.

Token estimation (1 token ≈ 4 chars) is an approximation. Inside ~15% of precise token counts for English prose [6], however misfires for code and non-Latin scripts. For manufacturing, swap in tiktoken [8] — it’s a one-line change in compressor.py.

The extractive compressor scores by query-token recall overlap: what number of question tokens seem within the sentence, as a fraction of the question size. That is quick and dependency-free however misses semantic similarity — a sentence that paraphrases the question with out sharing any tokens scores zero. Embedding-based sentence scoring would repair that at the price of an extra mannequin name per compression move.

Commerce-offs and What’s Lacking

Cross-encoder re-ranking. The _rerank() interface is already designed to be swapped out. Drop in a BERT-based cross-encoder for meaningfully higher pair-wise rankings.

Embedding-based compression. Substitute the token-overlap sentence scorer in _extractive() with a small embedding mannequin. Catches semantic relevance that key phrase overlap misses. In all probability value it for 100+ doc techniques.

Adaptive alpha. Classify the question sort dynamically and alter alpha relatively than utilizing a hard and fast 0.65. A brief question with uncommon area phrases most likely desires extra TF-IDF weight; a protracted natural-language query desires extra embedding weight.

Persistent reminiscence. The present Reminiscence class is in-process solely. A light-weight SQLite backend with the identical add() / get_weighted() interface would survive restarts and allow cross-session continuity.

Closing

RAG will get you the proper paperwork. Immediate engineering will get you the proper directions. Context engineering will get you the proper context.

Immediate engineering decides how the mannequin thinks. Context engineering decides what it will get to consider.

Most techniques optimise the previous and ignore the latter. That’s why they break.

The complete supply code with all seven demos is at: https://github.com/Emmimal/context-engine/

References

[1] Lewis, P., Perez, E., et al. (2020). Retrieval-Augmented Era for Data-Intensive NLP Duties. NeurIPS 33, 9459–9474. https://arxiv.org/abs/2005.11401

[2] Karpathy, A. (2025). Context Engineering. https://x.com/karpathy/status/1937902205765607626

[3] Pedregosa, F., et al. (2011). Scikit-learn: Machine Studying in Python. JMLR 12, 2825–2830. https://jmlr.org/papers/v12/pedregosa11a.html

[4] Baddeley, A. (2000). The episodic buffer: a brand new part of working reminiscence? Traits in Cognitive Sciences, 4(11), 417–423.

[5] Mihalcea, R., & Tarau, P. (2004). TextRank: Bringing Order into Texts. EMNLP 2004. https://aclanthology.org/W04-3252/

[6] OpenAI. (2023). Counting tokens with tiktoken. https://github.com/openai/tiktoken

[7] Nogueira, R., & Cho, Okay. (2019). Passage Re-ranking with BERT. arXiv:1901.04085. https://arxiv.org/abs/1901.04085

[8] OpenAI. (2023). tiktoken: Quick BPE tokeniser to be used with OpenAI’s fashions. https://github.com/openai/tiktoken

[9] Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings utilizing Siamese BERT-Networks. EMNLP 2019. https://arxiv.org/abs/1908.10084

Disclosure

All code on this article was written by me and is authentic work, developed and examined on Python 3.12.6. Benchmark numbers are from precise demo runs on my native machine (Home windows 11, CPU solely) and are reproducible by cloning the repository and operating demo.py, besides the place the article explicitly notes numbers are calculated from identified inputs. The sentence-transformers library is used as an elective dependency for embedding era in hybrid retrieval mode. All different performance runs on the Python commonplace library and numpy solely. I’ve no monetary relationship with any device, library, or firm talked about on this article.