recallia 0.0.2

recallia

Minimal voice-to-notes pipeline with FunASR + Ollama.

Quickstart (Docker)

1. Requirements: Docker & docker-compose; NVIDIA GPU + container toolkit recommended; Ollama (runs as a service in compose).
2. (Optional) Reuse host Ollama models: set HOST_OLLAMA_MODELS=/path/to/host/.ollama before docker compose up.
3. Build & run:

   docker compose up --build
   

4. Drop audio files (.mp3/.wav/.m4a/.aac) into data/inbox/. Outputs land in data/archive/YYYY-MM-DD/<hash>/.

Flow

• ASR via FunASR (GPU if available).
• Summaries/tags + per-sentence emotions via Ollama.
• Artifacts per file: original audio, transcript.jsonl, llm_summary.json, llm_emotion.json, run.log.
• Metadata in SQLite data/db.sqlite3.

Config

• docker-compose.yaml: OLLAMA_HOST (defaults to http://ollama:11434), GPU flags, caches (./cache, ./modelscope, optional HOST_OLLAMA_MODELS).
• recallia/config.py: ASR device, LLM model, base paths.

Local (no Docker)

uv sync
python -m recallia

Ensure Ollama reachable at OLLAMA_HOST.

Demo data

Use examples/generate_demo_data.py to produce sample audio/text for testing.

Example archive (real output)

data/archive/2025-12-15/7daf3d771c04/

- 02_emergency_call.mp3
- transcript.jsonl
- llm_summary.json
- llm_emotion.json
- run.log

• run.log key fields:
  • summary: 技术人员发现API网关出现502错误，原因是数据库连接延迟导致CPU负载过高和慢查询问题。决定手动添加索引并切换流量以缓解当前状况。
  • tags: api网关, 数据库连接, 索引
  • duration_sec: 7.064
• llm_summary.json snippet:

{
  "response": {
    "summary": "技术人员发现API网关出现502错误，原因是数据库连接延迟导致CPU负载过高和慢查询问题。决定手动添加索引并切换流量以缓解当前状况。",
    "tags": "api网关, 数据库连接, 索引"
  }
}

• transcript.jsonl starts with full text, then per-segment entries. Excerpt:

{"type": "full_text", "text": "Speaker 0: 喂，老张听得到吗？出大问题了， api 网关全是五百零二报错。\nSpeaker 2: 听到了，我刚看监控，数据库连接迟爆了， cpu 直接飙到百分之一百，\n..."}
{"type": "segment", "speaker": "Speaker 0", "text": "喂，", "start_time": 290, "end_time": 530, "emotion": "neutral"}
{"type": "segment", "speaker": "Speaker 2", "text": "数据库连接迟爆了，", "start_time": 10860, "end_time": 12400, "emotion": "neutral"}