| name | embedding-search |
| description | Design or implement semantic search, embeddings, vector indexes, RAG retrieval, document Q&A, chunking, similarity search, reranking, and retrieval evaluation. Use when information must be found by meaning rather than exact text. |
Embedding Search
Use this skill for retrieval systems over unstructured content. Use exact text search for literal matching, database-query for relational data, and ingestion skills to normalize source documents first.
Required contract
- Define the corpus, exclusions, sensitivity constraints, and expected query types.
- Normalize content and chunk on semantic boundaries. Start around 250–500 words with 10–20% overlap, then tune from evidence.
- Give every chunk a stable ID and source metadata suitable for citations and filtering.
- Choose one embedding model per index version. Record the model, dimensions, chunking policy, and index version; rebuild when those change.
- Use the same model for indexing and queries. Validate vector dimensions before writes or searches.
- Retrieve a bounded top-k set, optionally filter and rerank, deduplicate redundant passages, and preserve source citations.
- Require a grounded fallback when retrieved context is insufficient.
- Evaluate with a small gold query set before tuning prompts: retrieval hit quality, citation correctness, failure classes, latency, and cost.
Prefer a local-first, single-runtime implementation unless the user requests hosted infrastructure. Keep API keys outside code and pin request-specific dependencies.
Read the detailed guide for architecture choices, model/store tradeoffs, TypeScript structure, and tuning guidance.
