docmason
A repo-native agent application that transforms office documents into a locally-managed, traceable knowledge base. The core principle: every answer must be verifiable against original source documents. "The repo holds the truth. The agent does the reasoning."
Key Quotes
"The repo holds the truth. The agent does the reasoning."
"Enforces strict data contracts and provenance boundaries."
Key Themes
The problem docmason addresses is real: most document AI tools flatten complex files into unstructured text, losing the structural information that makes answers verifiable. Presentation layouts, spreadsheet relationships, formatting-based semantics, and cross-document connections all disappear. You get plausible answers with no way to check them.
docmason's approach: preserve original document structure, enforce strict source boundaries, and make every AI answer traceable to specific source material. It uses LibreOffice for high-fidelity parsing (important -- Office format conversion is notoriously lossy) and runs entirely locally on macOS.
The "one of many tools doing citations and source weighting" observation in the annotation captures a trend. As LLMs become more capable at analysis, the bottleneck shifts to trustworthiness. Provenance and citation become the differentiator.
Critical Analysis
Alpha status and a recommendation for "GPT 5.4 or equivalent" for reliable answers suggests this is early. The concept is strong but the execution may not be there yet.
The local-only approach (zero transmission of content) is the right call for corporate documents but limits the model quality you can use. Unless you're running a capable local model, you're sending content through whatever API your agent uses.
Compare with markitdown, which converts documents to flat Markdown (good for feeding into LLMs, bad for preserving structure), and graphify, which builds knowledge graphs from code (similar concept, different domain). docmason occupies the middle ground: structured knowledge extraction from office documents.
See also lat.md for another approach to validated, traceable knowledge bases.
Sources: docmason Last updated: 2026-05-14