Get Started

Framework

How Trunkbase turns a messy document into a clean, structured note — and hosts it for your AI, all on your machine. It does this by reading your documents the way a person would, recognizing layout, tables, and text, using machine learning that runs entirely on your own machine. No cloud, no uploads, no account required.

Documents go in as opaque files. They come out as clean, searchable, structured knowledge you fully own — with the intelligence of modern machine learning and the privacy of a local-first app.

text 
Decompose → Store → Find → Compose. Private by default.

The decomposition pipeline

Trunkbase uses a "smart parser first, AI where it's needed" approach. It doesn't throw AI at everything — it applies the right tool to each file, which keeps it fast, accurate, and private.

  • PDFs are parsed directly, page structure and all.
  • Office files (Word, PowerPoint, Excel) go through dedicated readers.
  • HTML is cleaned down to its real content.
  • Markdown and plain text pass straight through.

🧠 Neural layout understanding

For PDFs and images, an on-device layout model identifies the structure of a page — what's a heading, a paragraph, a caption, a column — instead of dumping text in reading-disorder. Your converted document keeps its shape.

📊 AI table recognition (TableFormer)

Tables are where most converters fall apart. Trunkbase uses TableFormer, a purpose-built machine-learning model that reconstructs a table's rows, columns, and merged cells — so a financial table in a PDF becomes a real, usable table, not a jumble of numbers.

🔎 Smart OCR — only when it's needed

Trunkbase is smart about when to use AI. It first checks whether a PDF already contains readable text:

  • Born-digital PDF (has a text layer) → read directly, OCR off. Fast and exact.
  • Scanned PDF or image (just pixels) → escalate to on-device OCR to recognize the text.

This "escalation-only" design means AI is used precisely where it adds value — never wastefully.

⚡ Deterministic where it's smart to be

Word, PowerPoint, Excel, HTML, and plain text are read by exact, deterministic parsers — no AI guesswork, because their structure is already known. Plain text passes straight through with no models at all.

Only the genuinely ambiguous parts escalate to local machine-learning models — a PyTorch layout model and a table-recognition model for tricky layouts, and RapidOCR for scanned pages and images. Deterministic parsers do the bulk; the models handle the hard 5%.

Under the hood

Decomposed notes are embedded and stored in a local LanceDB index, and search is hybrid — keyword (BM25) and vector similarity fused together, with an optional third signal from the knowledge graph, which is projected with networkx. Everything runs on-device; cloud LLMs are only ever an opt-in, bring-your-own-key extra, never a dependency.

Built with open source

Trunkbase stands on the shoulders of the open-source community. Our privacy promise is only credible because the technology underneath it is open and auditable — not a mystery cloud service. Here are the projects that make Trunkbase possible, with our thanks.

Document intelligence

  • Docling (IBM) — the document understanding engine behind our decomposer: neural layout analysis and the TableFormer table-recognition model.
  • PyTorch — the machine-learning runtime powering our on-device layout and table models.
  • RapidOCR — fast, on-device optical character recognition for scanned pages and images.

Search & knowledge

  • LanceDB — the embedded vector + keyword database powering hybrid semantic search.
  • FastEmbed (Qdrant) — efficient, local text embeddings.
  • NetworkX — the graph engine behind the Vault Knowledge Graph.

App foundation

Type & design