local-rag

Local RAG

Local-first semantic search: rag chunks a corpus, embeds it through the configured Ollama endpoint (default localhost), and indexes it with turbovec. A remote CONTEXT_KIT_OLLAMA_HOST receives corpus chunks and queries.

Prerequisites

• ollama serve running and the model pulled: ollama pull nomic-embed-text.
• Claude Code: the rag CLI is bootstrapped automatically on session start (uv venv). If rag is missing, run bash "${CLAUDE_PLUGIN_ROOT}/scripts/bootstrap.sh".
• GitHub Copilot/manual: clone this repo, set CONTEXT_KIT_DATA, run plugins/local-rag/scripts/bootstrap.sh, and add plugins/local-rag/bin to PATH.

Portable environment variables prefer CONTEXT_KIT_DATA, CONTEXT_KIT_EMBED_MODEL, and CONTEXT_KIT_OLLAMA_HOST; Claude plugin variables remain supported as fallbacks.

Use

rag index /path/to/vault --name notes      # build/update the index (incremental)
rag query "how did we handle retry backoff" --name notes --k 8
rag query "retry backoff" --name notes --k 8 --hybrid
rag status --name notes                     # counts, model, dim, FTS5 capability
rag list                                    # known indexes
rag remove --name notes --yes               # permanently remove one named index

Re-running index re-embeds only changed files (content hash). Results report path > heading + a snippet; JSON adds source offsets and retrieval-signal metadata. Follow up with rg to pin exact lines.

remove is non-interactive and refuses to run without --yes. It accepts the same safe index names as index, query, and status, removes only that named index's flat artifact directory, and fails clearly when the index is missing or cleanup is incomplete. These operations share a per-index process lock, so removal also refuses while the index is in use.

Hybrid retrieval (compose semantic + lexical modalities)

Semantic-only is the default. Add --hybrid to fuse turbovec and SQLite FTS5/BM25 candidates with deterministic RRF: 1.0 / (60 + semantic_rank) + 1.0 / (60 + lexical_rank). Each source retrieves 3 × k candidates before final fusion. rag status reports the fts5 capability; if unavailable, --hybrid fails clearly while semantic queries keep working.

Pipe a candidate file set into --allowlist - to limit both sources:

# From the obsidian bridge (graph/tags), or any tool that emits file paths:
obsidian backlinks file="Project X" | rag query "open risks" --name notes --hybrid --allowlist -
rg -l '#decision' "$VAULT" | rag query "why did we choose X" --name notes --hybrid --allowlist -

rag is not rtk-wrapped, so rtk rag … is a no-op (passes through). When rtk is installed, prefix the surrounding rg step instead — rtk rg -l keeps -l raw, so the piped paths above stay intact.

When NOT to use

For exact tokens/identifiers or code structure, prefer code-search (rg/sg) — it's faster and more precise. Reach for RAG when meaning ≠ words or the corpus is large/unfamiliar prose. See the retrieval-strategy skill to choose/compose.