# food-record-pipeline

> End-to-end food-record evaluation pipeline. Use when running VLM evals, CLIP food/non-food screening, CLIP→VLM batch extraction with resume, generating food calendars (month/week), or launching full background runs. Includes one-click pipeline runner and supporting scripts.

- Author: Daydreamer
- Repository: DayDreammy/food-record-pipeline-skill
- Version: 20260126000259
- Stars: 0
- Forks: 0
- Last Updated: 2026-02-06
- Source: https://github.com/DayDreammy/food-record-pipeline-skill
- Web: https://mule.run/skillshub/@@DayDreammy/food-record-pipeline-skill~food-record-pipeline:20260126000259

---

---
name: food-record-pipeline
description: End-to-end food-record evaluation pipeline. Use when running VLM evals, CLIP food/non-food screening, CLIP→VLM batch extraction with resume, generating food calendars (month/week), or launching full background runs. Includes one-click pipeline runner and supporting scripts.
---

# Food Record Pipeline

## Overview
Run VLM tests, CLIP screening, CLIP→VLM batch extraction with resume, and generate food calendars. Provides a one-click background pipeline script plus individual utilities.

## One-click pipeline
- **Full run (background):**
  - `./scripts/run_food_record_pipeline.sh`
  - Starts CLIP→VLM full pipeline with threshold 0.004 and logs to `eval/clip_to_vlm_full_<timestamp>.log`.

### Environment overrides (recommended for open-source use)
These scripts default to `ROOT_DIR=$(pwd)` and relative paths. Override as needed:
- `ROOT_DIR`, `PYTHON_BIN`
- `THUMBS_DIR`, `FULL_DIR`, `PROMPT_FILE`
- `CLIP_THRESHOLD`, `SEED`
- `PROXY_URL` (or set `HTTP_PROXY/HTTPS_PROXY/ALL_PROXY`)
- `OUT_DIR`, `LOG_FILE`

## Individual scripts
- **VLM eval (single dataset):** `scripts/run_vlm_eval.py`
- **CLIP food/non-food eval:** `scripts/clip_food_nofood_eval.py`
- **CLIP sample thumbs:** `scripts/clip_sample_thumbs_eval.py`
- **CLIP → VLM pipeline (resume + threshold):** `scripts/clip_to_vlm_pipeline.py`
- **Background full run (generic):** `scripts/run_clip_to_vlm_full_bg.sh`
- **Calendar (month/week):** `scripts/make_food_calendar.py`
  - Examples:
    - `./scripts/make_food_calendar.py --month 2025-12 --mode month`
    - `./scripts/make_food_calendar.py --month 2025-12 --mode week`
- **Calendar UI data builder:** `scripts/build_food_calendar_ui_data.py`
  - Generates `assets/food_calendar/data/weeks.js` from VLM jsonl.
  - Add `--use-vlm-analysis` to call VLM and fill weekly `summary/insights/advice`.
  - Uses `--analysis-cache` to reuse weekly analysis and `--refresh-analysis` to force refresh.
  - Add `--thumbs-dir` to prefer thumbnails in the UI (faster loading, better printing).
- **Export PDF (one-click):** `scripts/export_food_calendar_pdf.sh`
  - Input: CSV + thumbs + full images, outputs a printable weekly PDF.
  - Uses `scripts/print_food_calendar_pdf.py` (Playwright if installed, otherwise Chrome headless).

## Key defaults
- Proxy: `http://127.0.0.1:7890`
- CLIP threshold: `0.004` (conservative food filter)
- VLM prompt file: `prompts/food_record_v0_2.json`

## Notes
- The pipeline writes intermediate JSONL files and supports resume by skipping already processed records.
- Use `--clip-threshold` to adjust food filtering precision/recall tradeoff.

## Resources
- `scripts/` contains all runnable utilities listed above.
- `assets/food_calendar/` static weekly UI (open `index.html`).