12 Retrieval Systems 3 min read 592 words

Advanced Retrieval Patterns (Dec 2025)

Beyond the basics, production RAG systems use specialized patterns to handle complex query-document gaps. In late 2025, these patterns are the "Secret Sauce" of high-precision search.

retrievalragsearchdeep
01decomposition multi-query

Query Decomposition (Multi-Query)

Complex user queries are often "Compound Queries."

  • User: "Compare our Q3 vs Q4 revenue and explain the drop."

  • Decomposition:

    1. "Q3 Revenue 2025"
    2. "Q4 Revenue 2025"
    3. "Reasons for Q4 revenue variance"
  • Implementation: Use an LLM to generate these 3 sub-queries, search the DB for ALL of them, and aggregate the context.
02embeddings hyde

Hypothetical Document Embeddings (HyDE)

Queries are short; documents are long. This "Asymmetry" causes retrieval failure.

  • Pattern:

    1. Take the user query.
    2. Ask the LLM: "Write a 1-paragraph hypothetical answer to this."
    3. Embed the hypothetical answer instead of the query.
  • Why?: The hypothetical answer is in the same "Vector neighborhood" as the real documents, leading to much higher recall.
03anthropic pattern

Contextual Retrieval (The Anthropic Pattern)

standardized by Anthropic in late 2024, this pattern solves Context Dilution.

  • The Problem: A chunk might say "It costs $200," but without the header, we don't know "It" is a "Widget-X."
  • The Pattern: During ingestion, for every 300-token chunk, have an LLM write a 50-token context string (e.g., "This chunk is about the pricing for Widget-X in the North American market").
  • Benefit: Increases retrieval precision by 30-50% for fragmented data.
04document enrichment

Iterative Document Enrichment

Instead of just storing the raw document, we store "Enriched" meta-data.

  • Summary: Store a 1-paragraph summary of the document.
  • Q&A Generation: Generate 5 questions this document answers and embed those with the document.
  • Status: By late 2025, most high-end RAG systems embed "Questions" rather than "Answers" to match the user's query intent.
05reranking standard

In-Context Reranking (The 2025 Standard)

With 2M context windows, we use Rank-by-Context.

  1. Retrieve Top 100 docs.
  2. Put all 100 in the context window.
  3. Ask the model: "Read these 100 docs and identify the 5 most relevant. Then, use those 5 to answer."
  • Win: This utilizes the model's Long Context Reasoning to perform reranking without needing a separate Cross-Encoder model.
06questions

Interview Questions

Q: Why is HyDE (Hypothetical Document Embedding) risky for some applications?

Strong answer: HyDE relies on "Hallucinating" a baseline answer to find real data. If the user's query describes something non-existent or logically impossible, the LLM will still generate a hypothetical answer. This can pull in "Incorrect but Semantically Similar" data from the database, reinforcing the model's initial hallucination. In 2025, we mitigate this by using a Hybrid approach: retrieve once with the real query (Keyword) and once with the HyDE query, and use RRF to combine them.

Q: What is the "Asymmetric Retrieval" problem?

Strong answer: Asymmetric retrieval refers to the fact that user queries are usually short (3-10 words) while document chunks are long (300-500 words). These inhabit different statistical distributions in the vector space, leading to "Distance Bias." High-performance systems solve this using Asymmetric Encoders (one model for queries, one for docs) or Query Expansion (HyDE) to "inflate" the query into a document-like distribution.

07references

References

  • Gao et al. "Precise Zero-Shot Dense Retrieval without Relevance Labels" (HyDE, 2023/2024)
  • Anthropic. "The Contextual Retrieval Playbook" (2024)
  • LlamaIndex. "Query Transformation Cookbook" (2025)

Next: Agentic Systems

• summary · added by this rebuild

Key takeaways

01

Queries and documents are the wrong shape

A 3-10 word query and a 300-500 word chunk sit in different statistical distributions; HyDE inflates the query into a hypothetical answer so it lands near the real documents.

02

HyDE can reinforce a false premise

If the query describes something that does not exist the model still writes a hypothetical answer, pulling in similar-but-wrong data. Run the literal query alongside it and fuse with RRF.

03

Add context at ingestion, not query time

Writing a roughly 50-token context line for each 300-token chunk fixes the chunk saying it costs $200 with no idea what it is; the page claims 30-50% better precision.

04

Compound questions need splitting first

Compare Q3 against Q4 revenue and explain the drop becomes three separate searches whose results are aggregated, because one embedding cannot represent three distinct intents.

05

Embed the questions, not just the answers

Generating five questions a document answers and embedding those matches user query intent directly, which the page says most high-end RAG systems now do at ingestion.