# Content Extract MCP

Python MCP scaffold for article content extraction, structured summary validation, deterministic filtering, and Markdown sink output.

## Run

```bash
pip install -e .
summary-mcp
```

The server exposes five tools:

- `extract_url_content`
- `extract_item_content`
- `filter_summary_result`
- `run_freshrss_openclaw_pipeline`
- `generate_article_summaries`

Article-summary post-processing (separate LLM optional):

Daily knowledge-base defaults:

- `IMA_DAILY_KNOWLEDGE_BASE_ID` — default IMA knowledge base ID for daily single-article summaries
- `IMA_DAILY_KNOWLEDGE_BASE_NAME` — default IMA knowledge base name (expected: `daily`)
- Runtime upload logic should verify the configured target before upload; if unavailable, try resolving by name and create `daily` if needed

- `article-summary` MCP tool (operates on existing extracted payloads)
- `scripts/run_article_summaries.py` CLI helper

Validate an LLM summary result:

```bash
validate-llm-result outputs/reference/summary/result.json --extracted outputs/reference/extracted/read-flow-2026.extracted.json
```

Run the minimal extraction-to-summary loop:

```bash
python scripts/run_summary_loop.py ^
  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
  --prompt outputs/prompts/llm-summary-prompt.txt ^
  --output outputs/reference/summary/result.loop.json
```

Pull FreshRSS entries and map them into normalized `item` objects:

```bash
set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
set FRESHRSS_USERNAME=bot
set FRESHRSS_API_PASSWORD=your-api-password
python scripts/pull_freshrss_items.py --limit 5 --mark-read
```

By default the script excludes entries already tagged as `read`. Add `--include-read` if you want the full reading list.
When `--mark-read` is enabled, fetched entries are marked as read after the script finishes successfully.

The script writes:

- `outputs/freshrss/raw/freshrss.raw.json`
- `outputs/freshrss/items/freshrss.items.json`

Pull FreshRSS entries and run content extraction for each mapped item:

```bash
set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
set FRESHRSS_USERNAME=osiman
set FRESHRSS_API_PASSWORD=your-api-password
python scripts/run_freshrss_extract.py --limit 1 --mark-read
```

By default the script excludes entries already tagged as `read`. When `--mark-read` is enabled, only entries with successful extraction are marked as read.

The script writes:

- `outputs/freshrss/raw/freshrss.raw.json`
- `outputs/freshrss/items/freshrss.items.json`
- `outputs/freshrss/extracted/freshrss.extracted.json`

Run the full FreshRSS pipeline and mark items as read only after the final OpenClaw delivery payload is written:

```bash
set FRESHRSS_API_BASE_URL=http://127.0.0.1:8081/api/greader.php
set FRESHRSS_USERNAME=osiman
set FRESHRSS_API_PASSWORD=your-api-password
set LLM_API_URL=https://api.deepseek.com
set LLM_API_KEY=your-llm-api-key
set LLM_MODEL=deepseek-chat
python scripts/run_freshrss_pipeline.py --limit 5 --mark-read
```

If you want to apply your personal engineering and AI-agent interest profile during filtering, pass a context file:

```bash
python scripts/run_freshrss_pipeline.py ^
  --limit 5 ^
  --context configs/filter_context.personal.json ^
  --mark-read
```

This is the recommended production entrypoint. By default it writes only:

- `outputs/freshrss/rerun/<timestamp>/raw/freshrss.raw.json`
- `outputs/freshrss/rerun/<timestamp>/candidates/openclaw-delivery-payload.json`
- `outputs/freshrss/rerun/<timestamp>/run-report.json`
- `outputs/freshrss/rerun/<timestamp>/extracted/item-XX.extracted.json` (one per item)

It also updates the daily keyword index runtime data:

- `data/term_index/daily/YYYY-MM-DD.json`
- `data/term_index/term_stats.json`

The main pipeline does not emit a batch-level `freshrss.extracted.json` file by default.
If you need additional per-item intermediates such as normalized items, summaries, filter decisions, candidate records, or candidate inputs, add `--debug-artifacts`.

When OpenClaw is connected to the MCP server, it should call `run_freshrss_openclaw_pipeline` for the same behavior directly through MCP. The tool also supports `debug_artifacts=true` when deeper inspection is needed.

Run deterministic filter rules against a structured summary result:

```bash
python scripts/run_filter_rules.py ^
  --summary outputs/reference/summary/result.loop.json ^
  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
  --output outputs/reference/filter/filter-decision.json
```

You can optionally pass a context file to inject interest topics or source tags:

```bash
python scripts/run_filter_rules.py ^
  --summary outputs/reference/summary/result.loop.json ^
  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
  --context outputs/reference/filter/filter-context.json ^
  --output outputs/reference/filter/filter-decision.with-context.json
```

Rule engine details and rule authoring guidance live in:

- `docs/design/filter-rule-engine-design.md`
- `docs/design/filter-rule-engine-usage.md`

Write a filtered result into the Markdown sink:

```bash
python scripts/run_markdown_sink.py ^
  --summary outputs/reference/summary/result.loop.json ^
  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
  --filter outputs/reference/filter/filter-decision.json
```

The script writes markdown notes under `knowledge-base/`.

Build an internal `ArticleCandidateRecord` and a slim `OpenClawCandidateInput`:

```bash
python scripts/run_article_candidate.py ^
  --summary outputs/reference/summary/result.loop.json ^
  --extracted outputs/reference/extracted/read-flow-2026.extracted.json ^
  --filter outputs/reference/filter/filter-decision.json ^
  --section-hint tools_and_workflows
```

The script writes by default:

- `outputs/reference/candidates/article-candidate-record.json`
- `outputs/reference/candidates/openclaw-candidate-input.json`

Build a batch OpenClaw delivery payload:

```bash
python scripts/build_openclaw_delivery.py ^
  --input-dir outputs/freshrss/candidates/batch ^
  --sort-by-rank ^
  --date 2026-03-25
```

The script writes by default:

- `outputs/reference/candidates/openclaw-delivery-payload.json`

Output layout details live in `outputs/README.md`.

Keyword index defaults live in:

- `configs/term_aliases.json`
- `configs/term_stopwords.json`
- `configs/term_cleanup_policy.json`
- `configs/term_watchlist.json`
- `configs/term_change_log.json`

You can also rebuild the keyword index from an existing delivery payload:

```bash
python scripts/build_keyword_index.py ^
  --input outputs/reference/candidates/openclaw-delivery-payload.json
```

Runtime keyword data is stored under `data/term_index/`.

The keyword cleanup review skill lives in:

- `skills/keyword-cleanup-review/`

To build a review bundle for the LLM skill:

```bash
python skills/keyword-cleanup-review/scripts/build_review_bundle.py ^
  --days 7 ^
  --top 50 ^
  --output outputs/term_index/review/keyword-cleanup-bundle.json
```

The review bundle now also carries cleanup governance context:

- cleanup thresholds from `configs/term_cleanup_policy.json`
- the current watch list from `configs/term_watchlist.json`
- recent applied changes from `configs/term_change_log.json`

The skill only produces review inputs and suggestions. It does not modify `term_aliases`, `term_stopwords`, or `filter_context.personal.json` automatically.

To preview accepted suggestions before writing any config files:

```bash
python scripts/apply_term_suggestions.py ^
  --suggestions outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.json ^
  --accept-watch Cron Heartbeat Memory ^
  --dry-run
```

Remove `--dry-run` to write the accepted changes. The script can also apply accepted `alias`, `stopword`, and `interest keyword` suggestions through `--accept-alias`, `--accept-stopword`, and `--accept-interest`. Accepted watch terms are written into `configs/term_watchlist.json`, and every applied action is appended into `configs/term_change_log.json`.


## Article-summary LLM configuration

Set a dedicated model for post-processing summaries without affecting the main pipeline:

- `ARTICLE_SUMMARY_LLM_API_URL`
- `ARTICLE_SUMMARY_LLM_MODEL`
- `ARTICLE_SUMMARY_LLM_API_KEY`

If these are not set, the summarizer falls back to the main `LLM_*` / `OPENAI_*` settings used elsewhere.

Example (PowerShell style):

```bash
set ARTICLE_SUMMARY_LLM_API_URL=https://api.deepseek.com
set ARTICLE_SUMMARY_LLM_API_KEY=your-article-summary-key
set ARTICLE_SUMMARY_LLM_MODEL=deepseek-chat
```

Then run, for example, either via the CLI script:

```bash
python scripts/run_article_summaries.py ^
  --extracted outputs/freshrss/extracted/freshrss.extracted.json ^
  --ids 12345 67890 ^
  --output-dir outputs/freshrss/single_summaries
```

…or through the MCP server tool `generate_article_summaries` exposed by `summary_mcp.server`:

- `extracted_path` (string): path to a single-item extracted JSON (e.g. `outputs/freshrss/rerun/<run_id>/extracted/item-01.extracted.json`) or a batch extracted JSON containing a `results` array.
- `selected_ids` (array of strings): one or more `item_id` values to summarize. Pass an empty array to summarize all items in the file.
- `output_dir` (optional string): directory to write Markdown summaries. If omitted, summaries are written under `single_summaries/` next to the extracted file.
- `llm_api_key` / `llm_model` / `llm_api_url` (optional strings): overrides for article-summary LLM settings. If omitted, the tool falls back to `ARTICLE_SUMMARY_*` or main `LLM_*` env vars as described above.

The tool returns a JSON array of file paths for the generated Markdown summaries.

The article summary uses a dedicated prompt (`outputs/prompts/article-summary-prompt.txt`) that is completely independent from the daily digest prompt. It outputs a structured knowledge note in Chinese with sections: 核心结论、主要论点、关键方法 / 机制、重要细节、可复用启发、关键词、主题.