RAG Chunk Retriever

What this skill does

Queries a pre-built vector database index to retrieve the top-k document chunks most relevant to the input query. Handles embedding the query using the same model used during indexing, performing ANN search, and filtering results by similarity score threshold. Returns results sorted by relevance with full source attribution for citation in the final LLM response.

Inputs

• query — natural language question or search string
• index_config — connection config: {provider, collection_name, api_key, environment}
• top_k — number of chunks to return (default: 5)
• score_threshold — minimum similarity score to include a chunk (default: 0.7, range: 0–1)

Outputs

• chunks — list of {text, source_file, page, score, chunk_id} objects sorted by score descending
• sources — deduplicated list of source file paths that contributed chunks
• latency_ms — query latency in milliseconds (embedding + ANN search combined)

Supported vector databases

• Chroma — provider: "chroma", runs locally or via ChromaDB server
• Pinecone — provider: "pinecone", requires API key and environment
• Weaviate — provider: "weaviate", requires cluster URL and API key

Example

config = {
    "provider": "chroma",
    "collection_name": "product_docs",
    "embedding_model": "text-embedding-3-small"
}
result = rag_chunk_retriever(
    query="How do I reset my password?",
    index_config=config,
    top_k=3,
    score_threshold=0.75
)

{
  "chunks": [
    {"text": "To reset your password, click Forgot Password on the login page...", "source_file": "docs/account.md", "score": 0.92},
    {"text": "Password reset emails expire after 24 hours...", "source_file": "docs/security.md", "score": 0.81}
  ],
  "sources": ["docs/account.md", "docs/security.md"],
  "latency_ms": 43
}