cosmix-indexd — vector knowledge base + indexer

cosmix-indexd is the mesh's semantic memory: it embeds markdown into vectors, stores them in SQLite, and answers similarity search — so an agent can recall docs, journals, and skills by meaning rather than keyword. It auto-indexes workspaces on commit, keeping recall fresh without a manual step.

What it is

A long-running Rust daemon that runs a text-embedding model (nomic-embed-text-v1.5, via candle) in-process and keeps a vector store in SQLite. It is the retrieval backend behind the knowledge loop: the context_search an agent runs before a task, and the skills/docs/journal recall the mcp bridge exposes.

Content is chunked by markdown section, embedded, and stored with source metadata (_doc, _journal, _memory, _plan, _spec). Search embeds the query the same way and returns the nearest chunks by cosine distance.

What it does

Embed + store markdown chunks as L2-normalised vectors in SQLite.
Semantic search — nearest-neighbour recall over stored chunks, filtered by source type or metadata, top-k by distance.
Auto-index on commit — a git post-commit hook calls indexd.index_file per changed file, so a workspace re-indexes incrementally as it changes (never a full re-index).
Incremental single-file indexing — index_file splits, embeds, and upserts one file; a background mode enqueues it fire-and-forget.
Model lifecycle — loads the embedding model on demand and can unload to return RSS to the OS.

Running it

/opt/cosmix/bin/cosmix-indexd

Runs under systemd as cosmix-indexd.service (identity User=cosmix-indexd). Config loads from /etc/cosmix/indexd/config.conf.mix (system) or the XDG user path (the legacy .toml fallback was removed). The daemon also listens on a Unix socket (embed.sock, path from config) for a raw embedding API used by in-process consumers, alongside the Bus surface. The embedding model is fetched and cached via the Hugging Face hub cache on first use.

Interfaces

Bus service indexd — registered on the local broker.

Verb	Purpose
`indexd.search`	semantic search: `{query, limit, ...}` → ranked chunks
`indexd.index_file`	embed + store one file: `{path}`
`indexd.props.{list,get,watch}`	SPEC-12 property surface

Unix socket (embed.sock) — a local embedding API for direct consumers.

Example Bus call:

bus_call("indexd", "indexd.search", {"query": "mesh DNS zones", "limit": 5})

Where it fits

Depends on cosmix-lib-config, cosmix-lib-props-store (SPEC-12 state), and cosmix-lib-daemon, plus candle + a SQLite backend. It registers on the local cosmix-noded broker. The cosmix-mcp bridge is its main client — context_search, skills_*, and the *_feedback tools all route through indexd — and git hooks in indexed workspaces call it directly.