Knowledge Ingestion Service

Harness-neutral, local-first knowledge ingestion for Obsidian and other local retrieval
tools. It turns links, text, videos, screenshots, PDFs, and local files into structured
Markdown knowledge cards. Hermes, Codex, OpenClaw, and other MCP clients share the same
adapters and processing service.

Current release: 0.3.0. Web and file ingestion run cross-platform. WeChat Channels
downloading is an optional, experimental macOS integration that requires the desktop
WeChat client and a local TLS proxy.

How it works

Hermes / Codex / OpenClaw / CLI / Web / Telegram
                  |
              MCP / FastAPI
                  |
             Source Adapter
                  |
             ContentItem
                  |
       Cleaner / OCR / Whisper / AI
                  |
      Classifier / Tags / Knowledge Linker
                  |
          Obsidian Markdown + SQLite

Every source is normalized into a ContentItem. To add a platform, implement
SourceAdapter.detect() and SourceAdapter.fetch(), then register the adapter in
backend/adapters/registry.py.

Supported sources

Quick start

Python 3.11 or newer is required. Media processing requires ffmpeg; OCR requires
Tesseract.

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[media,browser,mcp,dev]"
playwright install chromium
cp config.example.yaml config.yaml
uvicorn backend.main:app --host 127.0.0.1 --port 8787

Open http://127.0.0.1:8787, or use the CLI:

.venv/bin/python scripts/ingest.py 'https://example.com/article'
.venv/bin/python scripts/ingest.py 'https://feeds.example.com/show.rss'
.venv/bin/python scripts/ingest.py 'https://vimeo.com/123456'
.venv/bin/python scripts/ingest.py '/absolute/path/file.pdf'
.venv/bin/python scripts/ingest.py 'A note to keep' --title 'Quick note'

The same pipeline is available through the API:

curl -X POST http://127.0.0.1:8787/api/ingest \
  -H 'content-type: application/json' \
  -d '{"url":"https://example.com/article"}'

Configure AI and Obsidian

The AI layer uses an OpenAI-compatible Chat Completions endpoint. AI is disabled by
default; without a model the system still creates a local fallback summary. Enable AI
for classification, visual understanding, and richer tags:

export OBSIDIAN_VAULT_DIR=/absolute/path/to/ObsidianVault
export AI_ENABLED=true
export OPENAI_BASE_URL=http://127.0.0.1:11434/v1
export OPENAI_API_KEY=''
export OPENAI_MODEL=qwen2.5:7b
export OPENAI_VISION_MODEL=your-vision-model

You can set the same values in config.yaml. The config file, .env, database,
browser login state, and downloaded media are ignored by Git.

Knowledge linking uses qmd when available. If qmd is not installed, it falls back
to lexical matching over the latest 1,000 Markdown notes in the Vault. Cards are written
to a temporary file and atomically replaced so an indexer never sees a partial note.

MCP tools and Harness clients

scripts/knowledge_mcp.py is a harness-neutral stdio MCP server. It exposes:

• knowledge_ingest: ingest a URL, local file, or text and wait for the knowledge card
  to finish.
• knowledge_get_job: inspect the current state of a known ingestion job.
• knowledge_list_capabilities: list supported sources and input types.
• knowledge_wechat_prepare: refresh the local WeChat Channels window only when the
  client connection needs recovery.

Each Harness launches the same server with a Python environment that includes the hermes
extra and the absolute path to scripts/knowledge_mcp.py. Client-specific Skills are in
clients/hermes, clients/codex, and clients/openclaw; they contain routing guidance,
not duplicate adapters. See clients/README.md for installation commands.

Docker

cp .env.example .env
docker compose up --build

Docker is suitable for web pages, files, OCR, transcription, and the AI pipeline. When
the workflow needs the macOS WeChat client, system proxy, or a GUI browser login, run the
backend directly on the host. Compose binds the service to 127.0.0.1:8787.

WeChat Channels security boundary

The Channels integration uses the separately maintained
ltaoo/wx_channels_download project.
Its license and security boundary are separate from this repository. This project does
not distribute its binary, root certificate, cookies, or WeChat login data. See
integrations/wechat-channels/README.md for installation, licensing, proxy, and macOS
permission details.

The downloader creates a local TLS proxy. Use only a trusted, checksum-verified build and
never expose the downloader or this service to a LAN. The MCP tool temporarily switches
the HTTP/HTTPS proxy for the task and restores the previous settings afterward. The
original video is deleted only after both the Obsidian note and SQLite record have been
written successfully.

Verification

pytest -q
ruff check backend scripts tests

The test suite covers Markdown formatting, task recovery, text end-to-end ingestion, X
context, the WeChat Channels adapter, video transcoding, post-write cleanup, and input
classification. Real platform pages and login sessions change over time, so production
deployments should still perform a separate end-to-end check for each platform they use.

Contributing and license

Read CONTRIBUTING.md and SECURITY.md before submitting a change. Original project code
is licensed under Apache-2.0. Optional third-party components remain under their own
licenses; see THIRD_PARTY_NOTICES.md.