Summary
A library that keeps code as code.
Most ingestion pipelines flatten a technical document into prose chunks and lose the part you wanted. Here the chunker preserves code blocks separately from narrative, and the type and tag filters on search let you ask for one without the other.
What it is
A document library you build: markdown, Python, OpenAPI and HTML files, live URLs and raw strings are chunked and embedded, then searched semantically, with an in-memory note store alongside.
What you get
- A markdown, Python, OpenAPI or HTML file ingested and chunked, or a string passed straight through with tags
- An HTML page ingested from a URL, with Playwright available for content that needs JavaScript
- Batch ingestion for several sources in one call
- A smart ingestion path that keeps only code, configuration, markdown structure and technical definitions, chunking code blocks separately from narrative
- Semantic search over what is stored, filterable by chunk type and by tag
- Context retrieval and the list of ingested sources
- Chunk-level maintenance — retype a chunk, delete one by ID, or remove a whole source
- Notes stored under a note:// URI, with a prompt that summarises them at brief or detailed length
Requirements
A Python environment for the server plus the embedding backend it writes into. The smart ingestion path calls a Gemini Flash model, so that route needs Google API access; the plain ingest tools do not.