검색기 및 맥락 압축

A vector store knows how to search, but agents talk to a retriever — a thin interface with one job: given a query, return relevant documents.

Any vector store exposes as_retriever() to produce one.

The k parameter sets how many documents to return. Too few misses context; too many wastes tokens and adds noise.

Instead of a fixed count, you can return only documents above a relevance score. This avoids forcing irrelevant chunks when nothing good matches.

MMR balances relevance with diversity, avoiding near-duplicate chunks. It is great when documents repeat similar text.

Documents carry metadata (source, date, category). You can filter retrieval to a subset, e.g. only the current product version.

Even relevant chunks often contain unrelated sentences. Sending that noise to the LLM dilutes the answer and burns tokens.

Contextual compression shrinks each retrieved document to only the parts that matter for the query.

This wrapper sits in front of a base retriever and post-processes its results with a compressor.

LLMChainExtractor asks the LLM to pull only the sentences from each document that are relevant to the query, discarding the rest before it reaches the final prompt.

LLM extraction costs tokens. EmbeddingsFilter is a faster, cheaper alternative that drops documents below a similarity threshold using embeddings only — no extra LLM call.

Combine steps in a DocumentCompressorPipeline: first a cheap embeddings filter, then LLM extraction on what survives. This keeps quality high while controlling cost.

Because a compression retriever has the same interface as any retriever, you swap it into your RAG chain without changing the rest of the pipeline. Cleaner context usually means better, cheaper answers.

You learned to tune and refine retrieval:

• Convert a store with as_retriever() and tune k
• Use score thresholds, MMR, and metadata filters
• Contextual compression removes noisy text
• EmbeddingsFilter is a cheap alternative to LLM extraction
• Chain compressors for quality plus efficiency

Better retrieval is often the biggest lever for RAG quality.