Structure-first reading
Parses headings, tables, and sections so the engine understands a document's shape — not just a blob of text.
Live
DOCNEST
Documents, understood.
The structure-first knowledge engine. DOCNEST reads a document by its structure — not by re-stuffing it into a prompt — and answers from what it already understands. Often with zero tokens, always on your machine.
In testing, DOCNEST returned zero-token answers on 80% of factual queries at 100% accuracy — roughly 92% fewer tokens than re-prompting the document each time.
How it works
Answer from structure, not from spend
Index a document once into a portable .udf, then ask freely.
Factual lookups are resolved from structure — no model call, no token bill.
$ pip install docnest-ai
>>> from docnest.reader import UDFIndex
>>> idx = UDFIndex.load("report.udf")
>>> r = idx.query("What was Q3 revenue?")
>>> r.answer # "Q3 revenue was $38M, up 22% YoY."
>>> r.tokens_used 0 # answered from structure, no LLM
What you get
Built to read, not re-read
Zero-token answers
Factual lookups resolve from structure with no model call — fast, deterministic, and free to ask again.
Portable .udf output
Indexing produces an open, single-file knowledge container you own and can move anywhere.
MCP server
Exposes your knowledge over the Model Context Protocol, so MCP-aware assistants can query it directly.
Local-first
Runs entirely on your hardware. Documents never leave your machine unless you choose to send them.
Clean Python API
A small, readable API — load, query, inspect — that drops into scripts, notebooks, and apps.
In the platform
DOCNEST works with
Bring it into your stack
Open source, MIT-licensed, and yours to run. Start in a notebook in minutes.