User Guide

Search pipeline

Query reformulation, document expansion, and contextual chunking — tuning the retrieval pipeline.

Query reformulation#

Query reformulation rewrites your query before retrieval to bridge vocabulary gaps between how users phrase questions and how answers are written in documents.

Strategy	How it works	Best for
HyDE	Generates a hypothetical answer, then searches for documents similar to it	Factual questions with predictable answer phrasing
Multi-Query	Generates 3 alternative phrasings and merges results	Ambiguous or domain-specific queries
Step-Back	Generates a broader version of the query for high-level context	Narrow questions that need broader context

Tip

Start with Multi-Query — it's the most generally useful strategy. Each reformulation adds one LLM call to the search pipeline.

Document expansion#

Document Expansion improves answer quality for queries targeting a specific document (policy number, ticket ID, case number). When 3 or more of the top 5 results come from the same document, the system pulls in all chunks from that document so the AI sees the full picture.

Enable/Disable — off by default. Turn on when your content includes document-specific queries.
Max Expanded Chunks — limits how many chunks are pulled per document (default 50).

When to use

Document expansion works best with document-specific queries like "What are the exclusions for policy ABC-12345?" or "Summarize ticket HELP-4521". For broad topic searches, standard chunk retrieval is usually sufficient.

Contextual chunking#

Contextual chunking prepends document-level context to each chunk before embedding. Without context, a chunk like "The deadline is March 15" is ambiguous — with context, the system knows it's about "Q1 budget review deadlines from the Finance team".

Mode	How it works	Cost
Rule-Based	Prepends document title, source name, and metadata to each chunk	Free — no API calls
LLM	Uses LLM to generate a contextual summary for each chunk	One LLM call per chunk

Tip

Rule-Based mode gives 80% of the benefit at zero cost. Use LLM mode only for content where chunks are frequently ambiguous without additional context.

Note

After changing the contextual chunking mode, re-crawl your data sources to apply the new strategy to existing documents.

Hybrid search balance#

The hybrid search balance controls the weight between semantic (vector) and keyword (full-text) search. A value between 0.0 (pure keyword) and 1.0 (pure semantic). The default 0.70 works well for mixed content.

Lower values (0.30–0.50) — when exact terms matter (product codes, legal clauses).
Higher values (0.70–0.90) — when meaning matters more than wording.