A multi-agent research tool in both Web App and CLI, which can:

• Search (no limit) in Google / Baidu / Xiaohongshu via opencli
• Search BOSS Zhipin jobs via opencli
• Follow up on BOSS Zhipin job chats via opencli
• Watch video in Bilibili / YouTube and summarize via opencli
• Customize any other cli as agent tool

uv run -m deepfind.cli "Help me do a research on Robotic Hands, give me a TLDR table" --num-agent 1

uv run -m deepfind.cli "Help me summarize nVidia press conference https://www.bilibili.com/video/BV1EhwmzsEqB - What's new?" --num-agent 1

uv run -m deepfind.cli "Help me summarize https://www.bilibili.com/video/BV1tew5zVEDf What Saining Xie point of view in the interview" --num-agent 1 --quiet

uv run -m deepfind.cli "Help me summarize this YouTube talk https://www.youtube.com/watch?v=dQw4w9WgXcQ" --num-agent 1

uv run -m deepfind.cli "How people think Elon Musk in Xiaohongshu?" --num-agent 2

uv run -m deepfind.cli "Help me search Agent jobs on Boss" --num-agent 1

uv run -m deepfind.cli "help me survey robotics arms industry" --num-agent 3 --long-report-mode --once

uv run -m deepfind.cli --list-tools

Install everything needed for local GPU mode and Bilibili ASR in one setup step:

uv sync --extra media --extra local-llm --extra-index-url https://mirrors.aliyun.com/pytorch-wheels/torch_stable.html

Install opencli for web search and video download:

npm install -g @jackwener/opencli

Pre-download the ASR model on Windows (PowerShell) to avoid first-run delay:

hf download Qwen/Qwen3-ASR-1.7B --repo-type model

uv tool install bilibili-cli
uv tool install xiaohongshu-cli
uv tool install twitter-cli
bili whoami
xhs whoami
twitter whoami

The CLI auto-loads .env from the repo root.

cp .env.example .env

Minimal .env:

QWEN_API_KEY=...
QWEN_MODEL_NAME=qwen3-max

Local Ollama GPU mode:

DEEPFIND_LOCAL_BASE_URL=http://127.0.0.1:11434/v1
DEEPFIND_LOCAL_MODEL=qwen3.5:9b
DEEPFIND_LOCAL_API_KEY=ollama

Optional:

QWEN_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
DEEPFIND_LOCAL_MODEL=Qwen/Qwen2.5-7B-Instruct
DEEPFIND_LOCAL_QUANTIZATION=4bit
OPENCLI_BIN=opencli
TWITTER_CLI_BIN=twitter
XHS_CLI_BIN=xhs
BILI_BIN=bili
ASR_MODEL=Qwen/Qwen3-ASR-1.7B
DEEPFIND_AUDIO_DIR=audio
DEEPFIND_TOOL_TIMEOUT=90
GOOGLE_NANO_BANANA_API_KEY=...
GOOGLE_NANO_BANANA_MODEL=gemini-3.1-flash-image-preview
DEEPFIND_IMAGE_DIR=tmp
DEEPFIND_IMAGE_SIZE=2K

uv run -m deepfind.cli "What's new in Xiaohongshu?" --num-agent 2
uv run -m deepfind.cli "Help me summarize video https://www.bilibili.com/video/BV1tew5zVEDf" --num-agent 1
uv run -m deepfind.cli "Help me summarize video https://www.youtube.com/watch?v=dQw4w9WgXcQ" --num-agent 1
uv run -m deepfind.cli "same query" --num-agent 2 --quiet
uv run -m deepfind.cli "same query" 

Flags:

• query: required unless using --list-tools
• --num-agent: 1..4
• --max-iter-per-agent: default 50
• --quiet: disable formatted progress output
• --once: always exit after the first answer
• --json: print structured research JSON
• --long-report-mode: switch the final lead response from a concise overview to a long-form report
• --list-tools: print built-in tool names/descriptions and exit
• --gpu: use the local Ollama model configured by DEEPFIND_LOCAL_BASE_URL and DEEPFIND_LOCAL_MODEL

To run with Ollama locally:

ollama serve
ollama pull qwen3.5:9b
uv run -m deepfind.cli "same query" --gpu

--long-report-mode is explicit only. It does not auto-detect benchmark tasks, so turn it on when you want a long-form benchmark-style report.

When deepfind runs in a real terminal, it now stays open after the first answer so you can continue the conversation with the same multi-agent/tool setup.

uv run -m deepfind.cli "first question"

Follow-up commands:

• Press Enter on a blank line to skip it.
• Type exit or quit to leave the session.
• Use --once to force the old one-shot behavior.
• If GOOGLE_NANO_BANANA_API_KEY is set, you can ask for an image after a summary and it will be saved under tmp/.
• You can also ask for one or more standalone HTML slides after a summary and they will be saved under tmp/.

uv run -m deepfind.cli "first question" --once

Example follow-up in chat mode:

Generate a 16:9 cover image from that summary and save it under tmp/
Generate 3 HTML slides from that summary and save them under tmp/

The repo also includes a local web chat UI under web/ with saved chats, live run activity, an iPhone-friendly mobile drawer, and a mode switch for Fast (1 agent) and Expert (4 agents). When a compatible NVIDIA GPU are available, the composer also shows a GPU/Cloud mode toogle.

Backend:

uv run deepfind-web --reload

Frontend:

cd web
npm install
npm run dev

Open the Vite URL during development, or run npm run build so deepfind-web can serve web/dist directly.

The production shell now includes a web manifest, touch icon, and iPhone Safari safe-area handling so it works cleanly in Safari and when added to the Home Screen.

The repo includes deployment scripts for a Raspberry Pi 5 running Ubuntu at david@192.168.0.205.

One-time SSH key bootstrap from your local machine:

./scripts/bootstrap_rpi_ssh_key.sh

Notes:

• The bootstrap script is now key-only. It verifies passwordless SSH access and, if access is missing, tells you which public key to install on the Pi.
• It uses your existing ~/.ssh/id_ed25519.pub or ~/.ssh/id_rsa.pub by default.
• Override PUBKEY_PATH, RPI_HOST, RPI_USER, or RPI_SSH_PORT if needed.

Deploy the app after SSH keys are working:

./scripts/deploy_rpi.sh

What the deploy script does:

• Syncs the repo to /home/david/apps/deepfind-cli
• Copies the local .env file to the Pi
• Installs uv if it is missing
• Runs uv sync --frozen, npm ci, and npm run build on the Pi
• Installs and restarts a user-level deepfind-web systemd service with systemctl --user

No-password deploy requirements:

• Passwordless SSH to the Pi must already work for david@<host>.
• The Pi already needs python3 3.11+, curl, tar, and systemd.
• deploy_rpi.sh will install a user-local Node.js 20.x automatically when node/npm are missing or too old.
• deploy_rpi.sh will also install @jackwener/opencli under the same user so web_search works after deploy.
• Boot persistence for the user service requires a one-time admin command on the Pi:

sudo loginctl enable-linger david

Default LAN URL after deployment:

http://192.168.0.205:8000

Useful checks on the Pi:

systemctl --user status deepfind-web
curl http://127.0.0.1:8000/api/health

• Lead planner can use tools before task split, usually with a web_search -> web_fetch flow on the most promising URLs.
• Sub-agents call local tools such as web_search, web_fetch, boss_search, boss_detail, boss_chatlist, boss_send, xhs_search_user, xhs_user, xhs_user_posts, xhs_read, twitter_search, twitter_read, bili_transcribe, and youtube_transcribe.
• Lead synthesis merges worker reports, fills gaps, and can do another web_search -> web_fetch pass when evidence is weak or conflicting.
• Lead final answer turns the synthesis into the user-facing response, and only uses asset tools for final image/slide requests.

bili_transcribe is available to sub-agents and accepts either a Bilibili video URL or a raw BV... ID. It returns transcript text only (no summary generation). If audio/transcripts/<BVID>.txt already exists, the tool reuses it and skips download + ASR transcription.

youtube_transcribe is available to sub-agents and accepts either a YouTube video URL or a raw video ID. It calls opencli youtube transcript and caches grouped transcript text under audio/transcripts/youtube/<VIDEO_ID>.txt.

Setup (WSL):

uv tool install bilibili-cli
bili status

Artifacts:

• Segments: audio/<BVID>/seg_*
• Transcript: audio/transcripts/<BVID>.txt

python3 -m unittest discover -s tests -v

Runtime pipeline for a normal research turn:

1. lead-plan
2. sub-1 .. sub-N
3. lead-synthesis
4. lead-final

When a follow-up only asks to reformat the previous answer in chat mode, the app can skip research and run lead-format directly.

When you run:

uv run -m deepfind.cli "same query" --json

the CLI prints a structured envelope with this shape:

{
  "version": "research.v1",
  "query": "user query",
  "lead": {
    "overview_md": "final markdown answer",
    "key_points": [
      {
        "text": "synthesized finding",
        "citation_ids": ["c1"],
        "confidence": "high"
      }
    ],
    "disagreements": [],
    "next_steps": []
  },
  "agents": [
    {
      "agent_id": "sub-1",
      "task": "assigned task",
      "summary": "worker summary",
      "claims": [
        {
          "text": "worker finding",
          "citation_ids": ["c1"],
          "confidence": "medium"
        }
      ],
      "gaps": []
    }
  ],
  "citations": [
    {
      "id": "r1",
      "dedup_id": "c1",
      "url": "https://example.com/report?utm_source=news",
      "title": "",
      "publisher": "",
      "source_agent": "sub-1",
      "source_section": "claim",
      "source_index": 1
    }
  ],
  "citations_dedup": [
    {
      "id": "c1",
      "canonical_url": "https://example.com/report",
      "url": "https://example.com/report?utm_source=news",
      "title": "",
      "publisher": ""
    }
  ],
  "meta": {
    "num_agents": 2,
    "max_iter_per_agent": 50,
    "generated_at": "2026-04-07T00:00:00Z"
  }
}

Notes:

• citations stores raw citation occurrences with provenance back to the agent and section that emitted them.
• citations_dedup stores canonicalized URLs with tracking parameters removed.
• Worker and synthesis prompts expect exact source URLs in citations whenever a claim is backed by a tool result.
• If a worker returns non-JSON text, DeepFind keeps the text as summary, leaves claims empty, and records gaps: ["non_json_output"].