Longlichi: a vision-language model for real-world paperwork

Longlichi is the vision-language model behind Hashtee's document products. This page is a working overview of what it does, why we built our stack around a VLM rather than a classical OCR pipeline, and where the research is heading.

The problem with reading paper as text

Classical OCR answers one question: what characters are on this page? That's the wrong question for most business paperwork. An invoice is not a wall of characters — it's a set of typed fields, a line-item table, a stamp that changes the document's meaning, and a handwritten correction in the margin that overrides the printed value.

A vision-language model answers a better question: what does this document say? Given a scan and a schema of what you need — invoice number, party names, line items with quantities — Longlichi returns typed, structured fields directly. There is no brittle chain of detect-then-recognise-then-parse stages to break when the layout shifts.

What it handles

Fields against a schema

Extraction is driven by a document profile — the fields you define, with types. The model reads against that contract, so output lands in your systems without a mapping layer.

Tables and line items

Multi-row line items come back as structured rows aligned to your columns — the part of documents that plain OCR most consistently mangles.

Handwriting, stamps, and mess

Shop-floor paper is not office paper. Handwritten registers, rubber stamps, smudges, skew, and hurried phone photos are the normal case we design for, not the edge case.

Described captures

A plain-language description of what you need — typed or spoken — is enough to define a capture with no schema at all. Useful for one-off documents, and the seed of a reusable profile.

Where it runs

Longlichi serves every Hashtee document surface: the Dodo scan appliance and app, the cloud extraction API, and the SDK paths (TypeScript and Python, plus a browser embed). The record shape is identical everywhere, so anything that reads one Longlichi record reads them all.

What we're working on

Current research directions, in plain terms:

Multi-page documents — carrying context across pages so a six-page purchase order extracts as one record, not six.
Confidence you can act on — per-field signals that are honest enough to drive review-only-what's-doubtful workflows.
Smaller, closer models — pushing more of the read toward the edge, so documents can be processed without leaving the building.

If any of this overlaps with a problem you have in production, we'd like to hear about it — the fastest way is a short call.