SourceRecall

SourceRecall is a self-hosted repository intelligence service for AI coding assistants. It indexes clean Git working trees into Qdrant, combines semantic and lexical retrieval, reranks candidates through a separate Jetson NLP service, and exposes the results through HTTP and Model Context Protocol (MCP).

It is designed for a small trusted-network deployment where repository data must remain under the operator's control.

Highlights

• Safe repository containment beneath one managed root

• Clean-commit indexing with file, chunk, model, and schema provenance

• Authoritative full re-indexing that removes obsolete Qdrant points

• Hybrid vector and ripgrep retrieval with chunk-aware reciprocal-rank fusion

• Optional cross-encoder reranking through Jetson NLP

• Thin stateless Streamable HTTP MCP integration with structured tool responses

• Durable SQLite job and index status

• Validated YAML configuration and structured rotating-file logs

• Single-worker, serialized indexing model with explicit resource bounds

Architecture

VS Code / Codex / Continue agent
    |
    | Streamable HTTP MCP
    v
SourceRecall MCP :8071
    |
    v
SourceRecall API :8070
    |                 |
    |                 +--> managed Git repositories
    |                      + ripgrep lexical retrieval
    |
    +--> Qdrant :6333
    +--> Jetson NLP :8091 (/v1/embeddings and /v1/rerank)

SourceRecall owns repository, path, commit, chunk, and retrieval semantics. Jetson NLP remains a stateless inference service that only knows text, vectors, and scores.

Repository layout

config/             Versioned configuration template
contrib/            Client-specific SourceRecall retrieval policies
deploy/systemd/     API and MCP systemd units
diags/              Deployment validation utilities
docs/               Architecture, API, deployment, and operations guides
src/source_recall/  Installable Python package
tests/              Automated tests

API at a glance

curl http://source-recall:8070/live

curl -X POST http://source-recall:8070/index \
  -H 'Content-Type: application/json' \
  -d '{"repository":"episode-tracker"}'

curl -X POST http://source-recall:8070/search \
  -H 'Content-Type: application/json' \
  -d '{
    "repository":"episode-tracker",
    "query":"where are TV episode credits normalized?",
    "limit":8
  }'

When security.api_token is configured, add Authorization: Bearer <token> to all non-health API requests.

Production paths

See Deployment for the complete installation procedure.

Documentation

• Architecture

• API reference

• Configuration

• Deployment

• Development

• MCP client setup and retrieval policies

• Operations

• Initial audit

The MCP guide includes native VS Code, the Codex IDE extension, and Continue configuration, plus the client-specific instruction files under contrib/ that teach each agent when to use SourceRecall instead of its live workspace tools.

Current scope

The initial release performs authoritative full indexing. Incremental Git indexing, webhook ingestion, syntax-aware chunking, symbol graphs, and reference analysis are planned extensions rather than claimed features. Indexing jobs are durable for status reporting but execute in one local process; this is not a distributed job queue.

The services do not terminate TLS, and the MCP endpoint does not implement client authentication. Keep them on a trusted private network or place them behind an authenticated reverse proxy. See Security.

License

SourceRecall is available under the MIT License.