RAG and LocalLore

Retrieval-augmented generation (RAG) gives a language model relevant external information at request time. Instead of retraining the model on private data, a retrieval system selects useful records and places them in the model's working context.

LocalLore is a focused RAG system for prior Claude Code sessions:

1. Ingest: discover and parse Claude Code JSONL records.
2. Represent: store message text for keyword search and encode it as vectors for semantic search.
3. Retrieve: rank candidate messages for the current query.
4. Augment: return excerpts and optional neighboring messages through MCP.
5. Generate: Claude uses that retrieved evidence while answering the user.

flowchart TB
    subgraph Offline indexing
        S[Session messages] --> P[Parse and normalize]
        P --> K[FTS5 terms]
        P --> E[Embedding vectors]
    end
    subgraph Request time
        Q[User question] --> R[Hybrid retrieval]
        K --> R
        E --> R
        R --> X[Relevant excerpts and context]
        X --> C[Claude response]
    end

LocalLore does not generate the final answer and it does not automatically add every old message to every prompt. Its job is narrower: make likely-relevant history cheaply discoverable, then let Claude request bounded context.

The retrieval step uses two complementary representations. Read Keyword and vector search to learn how inverted indexes and embedding vectors turn stored messages into searchable data.

Because both the corpus and embedding model remain local, the design trades hosted-service scale for privacy, predictable operation, and a small deployment surface.