clawsqlite-knowledge（爪式数据库知识）

clawsqlite-knowledge (OpenClaw Skill)

clawsqlite-knowledge is a knowledge base Skill built around the PyPI package clawsqlite.

It is a thin wrapper:

• it does not vendor the source code and does not git clone any repository;
• during installation, it installs clawsqlite>=1.0.2 (with a workspace-prefix fallback when the runtime env is not writable);
• during runtime, it operates the knowledge base only through the clawsqlite knowledge ... CLI.

Its main capabilities are grouped into three areas:

1. Ingestion

- ingest from a URL (together with an existing fetch tool such as clawfetch); - ingest from a piece of text, an idea, or an excerpt (marked as a local source).

1. Retrieval

- hybrid retrieval (LLM-aware query_refine/query_tags + FTS/vec with automatic downgrade) - show a full record by id (including full content).

1. Reporting (via underlying CLI)

- build interest clusters (summary/tag embeddings → interest topics) - generate periodic interest reports (Markdown + PNG, optional HTML/PDF) for the current knowledge base, based on previously built interest clusters.

Installation (performed by ClawHub / OpenClaw)

This skill is meant to be installed and run inside an OpenClaw/ClawHub runtime. It assumes:

• Python 3.10+ is available in the Skill runtime environment;
• the environment can access PyPI to install the clawsqlite package;
• the runtime model has access to a workspace directory where this skill lives

under skills/clawsqlite-knowledge.

Stage 1 — install the skill shell

Use the OpenClaw CLI to install the skill into your active workspace:

openclaw skills install clawsqlite-knowledge

This step downloads the skill package from ClawHub into:

~/.openclaw/workspace/skills/clawsqlite-knowledge

At this point the directory only contains:

• SKILL.md
• manifest.yaml
• bootstrap_deps.py
• run_clawknowledge.py
• README.md / README_zh.md

The clawsqlite PyPI package itself is not yet guaranteed to be installed.

Stage 2 — install or upgrade clawsqlite (PyPI, >=0.1.4)

The second stage is handled by the bootstrap script declared in manifest.yaml:

install:
  - id: clawsqlite_knowledge_bootstrap
    kind: python
    label: Install clawsqlite from PyPI
    script: bootstrap_deps.py

bootstrap_deps.py is intentionally small and auditable. In simplified form:

requirement = "clawsqlite>=1.0.2"
cmd = [sys.executable, "-m", "pip", "install", requirement]
proc = subprocess.run(cmd)
if proc.returncode != 0:
    prefix = _workspace_prefix()
    subprocess.run([
        sys.executable,
        "-m",
        "pip",
        "install",
        requirement,
        f"--prefix={prefix}",
    ])

Semantics:

• First, it tries to install clawsqlite>=1.0.2 into the default venv used for

the Skill runtime.

• If that fails (e.g. read-only venv), it falls back to a **workspace-local

prefix**:

  <workspace>/skills/clawsqlite-knowledge/.clawsqlite-venv

• On success in the prefix, it prints a NEXT: hint describing:

- Where the package is installed; and - Which site-packages path will be added to PYTHONPATH at runtime.

From the OpenClaw CLI, you typically do not need to call bootstrap_deps.py manually; openclaw skills install clawsqlite-knowledge (or a future openclaw skills update ...) will run the install hooks. If you want to force a re-install or upgrade of clawsqlite to 0.1.4+ inside the skill directory, you can run:

cd ~/.openclaw/workspace/skills/clawsqlite-knowledge
python bootstrap_deps.py

Where does the clawsqlite CLI live?

Depending on how pip is configured:

• If the first pip install succeeds in the base env, the clawsqlite

command and clawsqlite_cli module live in that venv;

• If we fall back to the workspace prefix, clawsqlite will be installed

under .clawsqlite-venv and the Skill runtime adds its site-packages directory to PYTHONPATH before invoking run_clawknowledge.py.

For advanced users, this means you can also invoke the CLI manually from the same prefix, for example:

cd ~/.openclaw/workspace/skills/clawsqlite-knowledge
PYTHONPATH="$(python - << 'EOF'
from bootstrap_deps import _workspace_prefix, _site_packages
p = _workspace_prefix()
print(_site_packages(p))
EOF)"$PYTHONPATH" \
  python -m clawsqlite_cli knowledge --help

In normal Skill usage (agents calling the JSON API), you do not need to manage this manually.

Runtime entry

The Skill runtime calls run_clawknowledge.py. This script:

• reads a JSON payload from stdin;
• routes by the action field to the matching handler;
• calls python -m clawsqlite_cli knowledge ... to perform the actual operation;
• writes the result JSON back to stdout.

All CLI calls are centralized in one function, which also injects the workspace-prefix site-packages path into PYTHONPATH when present so that the fallback installation works transparently.

If the underlying CLI emits NEXT: hints, this runtime surfaces them as a structured next array in the JSON response. On failure, it also includes an error_kind field for quick classification.

Supported actions

1. ingest_url

Ingest an article from a URL. The actual fetching logic is determined by the environment variable CLAWSQLITE_SCRAPE_CMD (recommended: the clawfetch CLI). This Skill does not fetch web pages directly.

Example payload:

{
  "action": "ingest_url",
  "url": "https://mp.weixin.qq.com/s/UzgKeQwWWoV4v884l_jcrg",
  "title": "WeChat article: Ground Station project",   // optional
  "category": "web",                                   // optional (default: web)
  "tags": "wechat,ground-station",                     // optional
  "gen_provider": "openclaw",                          // optional: openclaw|llm|off (default: openclaw)
  "root": "/data/clawsqlite-knowledge"                 // optional storage directory
}

Behavior:

• calls clawsqlite knowledge ingest --url ...;
• by default uses provider=openclaw:

- generates a long summary with heuristics (first ~800 characters, cut by sentence/paragraph boundaries); - generates tags with jieba or a lightweight algorithm (in 0.1.4 these are backed by the new keyword/semantic pipelines); - if embedding configuration is complete, generates an embedding for the long summary and stores it in the vec table;

• filenames use pinyin plus an English slug for easier cross-platform storage;
• the database keeps the original title and source_url.

Returns:

{
  "ok": true,
  "data": { "id": 1, "title": "...", "local_file_path": "...", ... }
}

2. ingest_text

Ingest a piece of text, an idea, or an excerpt, marked as a local source.

Example payload:

{
  "action": "ingest_text",
  "text": "Today I had an idea about a web scraping architecture...",
  "title": "Notes on web scraping architecture",   // optional; auto-generated if omitted
  "category": "idea",                              // optional (default: note)
  "tags": "crawler,architecture",                  // optional
  "gen_provider": "openclaw",                      // optional
  "root": "/data/clawsqlite-knowledge"             // optional storage directory
}

Behavior:

• calls clawsqlite knowledge ingest --text ...;
• generates long summary, tags, and embedding the same way as in the URL case,

depending on configuration;

• source_url will be Local;
• filenames use pinyin / English slug for easier cross-platform handling.

3. search

Search the knowledge base using the full clawsqlite>=1.0.2 search pipeline (query_refine/query_tags + FTS/vec hybrid), with automatic downgrade when embeddings or vec0 are not available.

Example payload:

{
  "action": "search",
  "query": "web scraping architecture",
  "mode": "hybrid",               // optional: hybrid|fts|vec (default: hybrid)
  "topk": 10,                     // optional
  "category": "idea",             // optional
  "tag": "crawler",               // optional
  "include_deleted": false,       // optional
  "root": "/data/clawsqlite-knowledge"   // optional storage directory
}

Behavior (high level):

• Calls clawsqlite knowledge search ... with --json and forwards filters.
• Uses the new four-mode capability model inside clawsqlite:

- Mode1: LLM + Embedding → query_refine + query_tags from SMALL_LLM, content/tag vectors + FTS + lexical tags. - Mode2: LLM + no Embedding → LLM-based query_refine/query_tags + FTS + lexical tags. - Mode3: no LLM + Embedding → heuristic query_refine/query_tags + content vectors + tag vectors + FTS/lexical tags. - Mode4: no LLM + no Embedding → heuristic query_refine/query_tags + FTS + lexical tags only.

• In all modes, natural-language queries are converted into:

- query_refine: a single, search-friendly sentence; - query_tags: a small set of keywords (length controlled by CLAWSQLITE_SEARCH_QUERY_TAG_MIN/MAX).

• When embeddings are available, the search scorer uses:

- content vectors (summary-based) for semantic similarity; - tag vectors + lexical tag matches as a tag channel; - FTS rank for BM25-like keyword matching; - priority and recency as light biases.

• Tag scoring is split into semantic and lexical parts, controlled by

CLAWSQLITE_TAG_VEC_FRACTION and CLAWSQLITE_TAG_FTS_LOG_ALPHA.

• Final scores are a weighted sum of these channels, with per-mode default

weights tunable via CLAWSQLITE_SCORE_WEIGHTS_MODE1..4 (and legacy CLAWSQLITE_SCORE_WEIGHTS*).

This skill does not re-implement scoring; it simply forwards the JSON result and lets agents inspect score, score_components, and any next hints surfaced by the underlying CLI.

Returns:

{
  "ok": true,
  "data": [
    {"id": 3, "title": "...", "category": "idea", "score": 0.92, ...},
    ...
  ]
}

4. show

Show one record from the knowledge base by id, optionally including full content.

Example payload:

{
  "action": "show",
  "id": 3,
  "full": true,                   // optional, default: true
  "root": "/data/clawsqlite-knowledge"   // optional storage directory
}

Behavior:

• calls clawsqlite knowledge show --id ... --full --json;
• returns full metadata and optional body content (the content field).

FTS/jieba fallback (CJK)

This Skill relies on the underlying clawsqlite CLI for FTS tokenization. When the CJK tokenizer extension libsimple cannot be loaded, clawsqlite can switch to a jieba-based pre-segmentation mode controlled by CLAWSQLITE_FTS_JIEBA=auto|on|off:

• auto (default): only enable when libsimple is unavailable and jieba is installed.
• on: force jieba pre-segmentation even if libsimple is available.
• off: disable jieba pre-segmentation.

In jieba mode, CJK text is segmented with jieba and joined with spaces before being written to the FTS index; queries apply the same normalization, so write/rebuild/query stay consistent.

If you change this setting on an existing DB, rebuild the FTS index:

clawsqlite knowledge reindex --rebuild --fts

Maintenance (CLI only)

This skill intentionally does not expose destructive maintenance actions via its JSON API. To clean up orphan files, old backups, or compact the knowledge database, use the clawsqlite CLI directly from a trusted administrative context, for example:

# Preview maintenance (no deletions)
clawsqlite knowledge maintenance prune \
  --root /data/clawsqlite-knowledge \
  --days 3 \
  --dry-run \
  --json

# Apply maintenance (delete orphans + old backups, then VACUUM)
clawsqlite knowledge maintenance prune \
  --root /data/clawsqlite-knowledge \
  --days 7 \
  --json

Only administrators or scheduled automation should run these commands. Agents using the clawsqlite-knowledge skill have access only to ingestion, retrieval, and show operations.

Security and auditability

• The Skill depends only on the clawsqlite package from PyPI.
• It does not vendor source code, does not git clone, and does not download

extra binaries.

• All knowledge base operations are performed through explicit

clawsqlite knowledge ... CLI calls, and their stdout/stderr can be fully audited in logs.