| Technique | Core Idea (1-liner) | Problem It Solves | When It Acts in RAG | Strengths | Limitations | Best Use Cases |
|---|---|---|---|---|---|---|
| HyDE (Hypothetical Document Embeddings) | Embed a hypothetical answer instead of the query | Poor / vague query embeddings | Before retrieval | Dramatically improves recall for vague queries; aligns with document style | Depends on LLM quality; extra cost | Research RAG, medical/legal QA, exploratory questions |
| Window Search | Retrieve neighboring chunks around a hit | Loss of context due to chunking | After retrieval (context expansion) | Preserves narrative flow; improves coherence | Increases token usage | PDFs, books, policies, manuals |
| Self-Query Retriever | LLM converts natural language into semantic query + filters | Hidden user constraints (time, level, type) | Before retrieval (query planning) | Powerful structured + unstructured search; enterprise-friendly | Needs clean metadata; schema-dependent | Product search, course catalogs, enterprise docs |
| Contextual Compression Retrieval | Shrinks documents to only query-relevant parts | Token waste & noisy context | After retrieval, before generation | Saves tokens; reduces hallucination; improves precision | Risk of over-compression; added compute | Long docs, cost-sensitive RAG, strict token limits |
| RAG Fusion (Multi-Query Retrieval) | Generate multiple query variants and merge results | Narrow recall due to wording bias | During retrieval (multi-pass) | High recall; uncovers blind spots | Higher latency & cost; needs reranking | Open-domain QA, research, knowledge-heavy RAG |
Continue With Google 
Continue With Facebook 
AI: Artificial intelligence
