RAG パイプラインは役に立たない可能性が高い。より良い代替案とは

**

# Introduction

Retrieval-augmented generation** (RAG) emerged as the standard approach for connecting documents with large language models (LLMs).

The pattern is simple: embed a corpus, retrieve the most relevant chunks by vector similarity, inject them into a prompt. It works well in demos and many production systems. It also fails in predictable, documented ways that only show up at scale.

Here is what those failure modes look like, and the alternatives engineers are reaching for to address them.

**

# When RAG Fails in Production

The most common failure pattern is retrieval irrelevance. A user queries a parental leave policy. The retriever returns the 2022 version, the 2024 version, and a cultural blog post. Each chunk scores high on embedding distance because it shares vocabulary with the query. None of them answers the question the user actually asked.

The model does not know the retrieved content is outdated or off-topic. It blends the chunks into a confident, detailed answer that is factually wrong. This is topical similarity without factual relevance, and it is the dominant failure mode in production RAG systems.

A subtler version is context poisoning. Enterprise knowledge bases often hold the same policy document in multiple versions. When the retriever returns chunks from both, the model does not surface the contradiction. It picks one, blends both, or presents a confident synthesis. The reader gets an answer. The answer may be wrong. Neither the user nor the model knows it.

The underlying cause is a structural conflict in the chunk-embed-retrieve pipeline. Good recall needs small chunks, around 100 to 256 tokens, for focused retrieval. Good context understanding needs large chunks, 1,024 tokens or more, for coherence. Every RAG designer picks one and accepts the trade-off.

# The Common (Wrong) Fix: Over-Engineering

When standard RAG underperforms, the common fix is to make it more complicated: higher-dimensional embeddings, more sophisticated reranking, multi-step retrieval. This compounds the problem.

A global manufacturing company budgeted \$400K for its RAG system. Year one cost \$1.2M. Final accuracy on technical documentation queries: 23%. The project was terminated. A healthcare enterprise hit \$75K per month in vector database costs by month six. These outcomes reflect a broader pattern: enterprise RAG implementations had a 72% first-year failure rate in 2025.

Higher embedding dimensions and more sophisticated vector models do not automatically improve performance. They raise compute costs and delay the more useful question, which is whether the retrieval architecture was the right choice at all.

# Alternatives When RAG Fails

// Long-Context Prompting

The most direct alternative to over-engineering a struggling RAG pipeline is to skip retrieval entirely.

If the corpus fits in the model's context window, load it and let the model read. A benchmark study found that long-context LLMs consistently outperformed RAG on QA tasks when compute was available, with chunk-based retrieval lagging the most.

The cost trade-off is significant. At 1M tokens, latency runs 30 to 60 times slower than a RAG pipeline, at roughly 1,250 times the per-query cost. With prompt caching for high-traffic applications, long-context can become cost-competitive.

A common decision rule: if the corpus fits in the context window and the query volume is moderate, long-context prompting is the cleaner starting point. Add retrieval only when the corpus exceeds the window, latency violates service level objectives (SLOs), or query volume crosses the economic break-even point.

// Memory Compression

When the corpus is too large for the context window, summarize before retrieving. Summarization-based retrieval compresses documents before injecting them, rather than pulling raw chunks. Benchmarks show this approach performs comparably to full long-context methods, while chunk-based retrieval consistently lags behind both.

One concrete result: an order-preserving RAG approach using 48K well-chosen tokens outperformed full-context retrieval at 117K tokens by 13 F1 points, at one-seventh the token budget. A well-compressed relevant document beats a raw dump of tangentially related chunks.

// Structured Retrieval

When retrieval is the right architecture, the solution is routing by query type rather than applying better embeddings uniformly.

Research from EMNLP 2024 introduced Self-Route, which lets the model classify whether a query needs full context or focused retrieval before running it. Simple factual lookups go to focused RAG. Complex multi-hop questions requiring global understanding go to a long context.

The result: better overall accuracy at a lower computational cost. Adaptive systems using this hybrid approach have shown 15 to 30% retrieval precision improvements through hybrid search and reranking.

The key change is making routing explicit. Every query gets classified before any retrieval runs, and the system stops treating all queries as identical embedding problems.

// Graph-Based Reasoning

For queries that require understanding relationships across a dataset rather than fetching a specific passage, vector retrieval fails by design.

These are the multi-hop questions: which decisions did the board reverse in Q3, and what was the stated reason each time? No single chunk answers this. The answer lives in the connections between documents.

Microsoft Research introduced GraphRAG** in 2024. The system builds a knowledge graph from the corpus, then traverses entity relationships rather than matching vectors.

**

It directly addresses the failure case that standard RAG cannot handle: synthesis across multiple documents requiring relational reasoning.

The trade-off is cost. Knowledge graph extraction runs 3 to 5 times more expensive than baseline RAG and requires domain-specific tuning. GraphRAG is worth the overhead for thematic analysis and multi-hop reasoning. For single-passage factual lookups, it is not.

# Conclusion

RAG is a reasonable default for many use cases.

It also breaks in predictable ways: retrieval irrelevance when vocabulary matches but semantics diverge, context poisoning when contradictory versions exist in the corpus, and structural limits when chunk size cannot satisfy both recall and coherence at once. Adding complexity to a broken retrieval design makes those problems more expensive.

There are four better paths, depending on the situation:

• If the corpus fits the context window, long-context prompting avoids the retrieval problem entirely.
• If context compression is necessary, summarization before retrieval outperforms raw chunk retrieval.
• If queries vary by type, explicit routing with structured retrieval improves both accuracy and cost.
• If queries require relational synthesis across documents, graph-based reasoning is the right architecture.

Match the architecture to the query type.

Nate Rosidi** is a data scientist and in product strategy. He's also an adjunct professor teaching analytics, and is the founder of StrataScratch, a platform helping data scientists prepare for their interviews with real interview questions from top companies. Nate writes on the latest trends in the career market, gives interview advice, shares data science projects, and covers everything SQL.