Retrieval Interfaces

components defines RAG contracts; it does not ship a complete vector database, retrieval engine, document splitter, or embedding service.

Contracts

Interface	Responsibility
Loader.Load	Load documents from a source
Embedder.Embed	Convert texts into corresponding vectors
Indexer.Index	Persist documents into a searchable index
Retriever.Retrieve	Retrieve documents for a query

Document carries ID, content, and metadata. Source is an implementation-defined identifier. The embedding contract requires one vector per supplied text in matching order; your implementation must also keep dimensions and model versions consistent with the index.

Compose a Retrieval Tool

retrieval := &agentcore.Tool{
    Name: "retrieve", Description: "Find relevant approved documents",
    Parameters: map[string]any{
        "type": "object",
        "properties": map[string]any{"query": map[string]any{"type": "string"}},
        "required": []string{"query"},
    },
    Func: func(ctx context.Context, raw json.RawMessage) (any, error) {
        var input struct { Query string `json:"query"` }
        if err := json.Unmarshal(raw, &input); err != nil {
            return nil, err
        }
        return retriever.Retrieve(ctx, input.Query)
    },
}

Here retriever is your components.Retriever implementation. Limit result count, content size, and allowed sources there. Register the tool with an agent, or compose retrieval and answering as explicit workflow stages when that order must be deterministic.

Keep Data Policy Outside the Prompt

Enforce tenant and document access before results reach model context. Retrieved content can contain malicious instructions; keep it separate from system policy. Preserve document IDs and provenance so answers can be audited and sources can be cited.

Design indexing, deletion, refresh, and embedding-version migration as host operations. Context compaction is not a substitute for a retrieval index, and the embedding tool selector ranks tools rather than providing document RAG.