Overview

This skill implements advanced RAG query transformation, combining query expansion, HyDE, and multi-query generation to boost retrieval relevance. By decomposing complex prompts, generating higher-level prompts, and routing results to specialized indices, it helps users find accurate information faster.

How This Skill Works

It integrates five transformation techniques: Multi-Query Generation to create query variations; HyDE to generate a hypothetical answer and its embedding; Query Decomposition to break complex requests into sub-queries; Step-Back Prompting to surface higher-level questions; and Query Expansion with synonyms and related terms. Configuration options include the number of variations, the LLM for generation, decomposition depth, query routing rules, and result fusion strategy.

When to Use It

• Queries are ambiguous or broad and require refinement
• Retrieval quality is insufficient or inconsistent
• Working with specialized knowledge bases that need routing to dedicated indices
• Queries involve multi-step reasoning or nested topics
• You need to balance latency and quality by tuning variation counts and fusion strategies

Quick Start

1. Step 1: Enable rag-query-transformation in your retriever and configure core options (number of variations, decomposition depth, routing rules).
2. Step 2: Enable Multi-Query Generation and HyDE, and set up result fusion to combine matches from multiple variants.
3. Step 3: Run representative queries, evaluate results for accuracy and latency, and iterate on routing and fusion settings.

Best Practices

• Match technique to query complexity to avoid over-processing
• Test with representative queries and monitor changes in results
• Tune the number of query variations and decomposition depth for your data
• Set clear routing rules to direct queries to appropriate indices
• Evaluate latency vs. quality tradeoffs and adjust fusion strategies

Example Use Cases

• Customer support knowledge base: use HyDE to generate a hypothetical answer and retrieve related policy docs for ambiguous questions like 'refund policy'.
• Enterprise policy search: combine multi-query generation with decomposition to retrieve relevant clauses across documents.
• Technical docs: decompose complex feature requests into sub-queries to locate precise specs.
• Specialized indices: route queries to finance or legal indices to improve relevance.
• Scientific literature: use step-back prompting to get higher-level summaries before diving into methods.

Frequently Asked Questions