JKinco Listen · 筑听（开源本地版）

Architecture

中文版 / Chinese version

Overview

flowchart LR
    U[Browser · React + TypeScript] -->|upload / live recording| API[FastAPI backend]
    API --> ASR[FunASR local recognition]
    ASR --> LEX[Domain lexicon correction]
    LEX --> CLS[Scene detection<br/>evidence gate + model review]
    CLS --> LLM[Local LLM via Ollama]
    LLM --> MIN[Structured minutes + quality review]
    MIN --> EXP[DOCX / PDF export]
    MIN --> DB[(SQLite local history)]
    DB --> QA[Ask JKinco · local meeting Q&A]

Modules

Module	Responsibility
frontend/	React + TypeScript workbench: upload, live recording, human review, export, history, meetings
backend/main.py	FastAPI routes, authentication, rate limiting, job queue, static assets
backend/meetings.py	Real-time meeting lifecycle (optional LiveKit), history archival
backend/auth.py	Local accounts, CAPTCHA, signed sessions, guest quotas
jkinco_asr.py	FunASR local speech recognition, duration gate, domain correction
jkinco_llm.py	OpenAI-compatible calls, primary/fallback model degradation, retries
jkinco_classifier.py	Scene detection: rule-based evidence gate plus model review
jkinco_reports.py	Chunked extraction, minutes synthesis, quality review, meeting overview
jkinco_export.py	DOCX / PDF export with original layout
jkinco_history.py	History storage, visibility rules, search
jkinco_assistant.py	Local Q&A over the current meeting plus history retrieval
jkinco_dingtalk.py	Optional DingTalk bot delivery with request signing

Data flow

1. The user uploads a recording, or records live in the browser.
2. ffmpeg probes the duration; oversized inputs are rejected before decoding.
3. FunASR transcribes locally, batching at 300-second intervals.
4. The domain lexicon corrects industry terms, personal names, and organization names.
5. The rule-based evidence gate classifies the scene first. The model only reviews that decision — it cannot override strong evidence.
6. The local LLM generates structured minutes, followed by a second pass checking facts, template conformance, and whether every action item has an owner.
7. Results land in local SQLite, ready for human review, export, and Q&A.

Why the evidence gate comes first

Scene detection decides which document template a meeting gets, and the wrong template changes what the minutes assert about who is accountable. A language model asked “is this a construction site review?” will say yes for any meeting that mentions projects, schedules, and quality — words that appear in almost every business meeting.

So the rules run first and they win. The model can downgrade a classification when evidence is thin, but it cannot promote one. This is a deliberate asymmetry: a meeting wrongly filed as “general minutes” is a formatting annoyance, while a meeting wrongly filed as a construction review is a document that misstates legal responsibility.

Security perimeter

• All inference happens in-process or against a local Ollama instance. Audio and text never leave the machine.
• No cloud credentials exist anywhere. CI rejects cloud-model traces and secrets.
• Login rate limiting, CAPTCHA, HMAC-signed sessions, CSP and security headers.
• Template uploads are hardened against zip bombs, path traversal, and XML entity attacks.
• Job slots are quota-limited separately for accounts and guests.

Testing

tests-v2/ contains 961 tests covering the API, security, concurrency, templates, export, scene rules, and module contracts. scripts/smoke_test.py provides an offline smoke entry point that exercises scene routing and every export path without needing a model.

This site is open source. Improve this page.