Embeddings & Search

When configured, q15 can build and search embedding collections over your own content. This is powered by Qdrant for vector storage and Gemini for embeddings.

How it works

1. Define sources — declare typed embedding sources under /workspace/.q15/embed/sources.json. Each source specifies a collection, source_type, and path.
2. Sync — embed_sync indexes the source content into Qdrant, storing both dense Gemini vectors and Qdrant-generated BM25 sparse vectors.
3. Search — embed_search runs hybrid dense + sparse search by default, combining semantic recall with lexical precision.

Collections

Collection	Purpose
library	Books, articles, and documents from your personal library
zettelkasten	Atomic knowledge notes from the zettelkasten
semantic	Extracted semantic memory content
core	Core identity and reference material

Source types

• markdown_tree — scans a directory tree of markdown files
• markdown_file — ingests a single markdown file
• chunked_markdown_tree — for pre-chunked markdown corpora (e.g., /workspace/library/<author>/<work>/chunks). Each markdown file is one embedding unit, and sibling meta.yml files provide metadata.

Scanner behavior comes from source_type, never from the collection name.

Configuration

Embeddings are optional. To enable them, add to the agent config:

agent:
  tools:
    embeddings:
      qdrant_url_env: Q15_QDRANT_URL
      gemini_api_key_env: Q15_GEMINI_API_KEY
      model: gemini-embedding-2
      dimensions: 768

Omit the embeddings block entirely to disable all embedding tools.

Search modes

• Hybrid (default) — dense Gemini vectors + Qdrant BM25 sparse vectors
• Dense — Gemini-only semantic search
• Sparse — BM25 lexical search only (no embedding call needed)