Aller au contenu principal

Task: Audierne Docs Processing

File: app/services/tasks/task_audierne_docs.py Schedule: 0 */2 * * * (Every 2 hours - dev mode) LLM Provider: Ollama (default) with gemini failover Resource Lock: lock:ollama:global (TTL: 600s)

Purpose​

Processes audierne2026 markdown documents to build training datasets for Forseti prompt optimization. Designed for continuous improvement of charter validation.

What It Does​

  1. Reads Progress File

    • Checks docs/docs/audierne2026/PROCESSING_PROGRESS.md
    • Identifies next unprocessed document

  2. Acquires Ollama Lock (if using Ollama)

    • Prevents concurrent Ollama usage with other tasks
    • TTL: 600 seconds (10 minutes)
    • Falls back to gemini if lock unavailable

  3. Processes One Document

    • Extracts themes using LLM (chunked for large docs: 15k chars, 500 overlap)
    • Generates contributions (2 per theme by default)
    • Includes violation examples for testing

  4. Creates Opik Dataset

    • Dataset name: audierne-{filename}-{date}
    • Contains input/output pairs for optimization

  5. Updates Progress

    • Marks document as complete
    • Logs themes, contributions, dataset name

  6. Releases Ollama Lock

    • Always releases in finally block

Documents​

Located in docs/docs/audierne2026/:

FileSizeDescription
Mayor-2026-wishes.md21KBMayor's New Year speech
municipal-council-january2026.md22KBCouncil meeting minutes
MVP_meeting.md13KBMVP planning meeting
MVP-meeting-satellite.md11KBSatellite meeting notes
overview.md6KBProject overview

Progress Tracking​

Progress file: docs/docs/audierne2026/PROCESSING_PROGRESS.md

| File | Status | Processed At | Themes | Contributions | Dataset |
|------|--------|--------------|--------|---------------|---------|
| MVP-meeting-satellite.md | ✅ Done | 2026-02-06 11:29 | 3 | 12 | audierne-MVP-meeting-satellite-20260206 |
| MVP_meeting.md | ⏳ Pending | - | - | - | - |

Configuration​

# Default settings in task file
DEFAULT_PROVIDER = "ollama"
FALLBACK_PROVIDER = "gemini"     # Used when Ollama unavailable
CONTRIBUTIONS_PER_THEME = 2
INCLUDE_VIOLATIONS = True
OLLAMA_LOCK_TTL = 600            # 10 minutes

Failover Behavior​

When Ollama is unavailable or locked by another task:

  1. Task checks lock:ollama:global in Redis
  2. If locked → automatically uses gemini instead
  3. Result includes provider_used to show actual provider
# Task result with failover
{
    "status": "success",
    "provider_used": "gemini",  # Failover occurred
    "warnings": ["Ollama busy, using gemini"]
}

Disable Failover:

# Force Ollama only (will skip if locked)
result = task_audierne_docs(
    provider="ollama",
    enable_failover=False
)
# Returns status="skipped", reason="ollama_locked" if busy

Manual Run​

# Run next pending document
python -c "from app.services.scheduler import run_task_now; print(run_task_now('task_audierne_docs'))"

# Force specific document
python -c "
from app.services.tasks.task_audierne_docs import task_audierne_docs
result = task_audierne_docs(force_doc='Mayor-2026-wishes.md')
print(result)
"

Success Keys​

Note: Uses skip_success_check=True to allow multiple runs per day (one doc at a time).

success:task_audierne_docs:YYYYMMDD  # Set when ALL docs are processed

Output​

Each run produces:

  • Opik Dataset: audierne-{filename}-{date} with training data
  • Progress Update: Markdown table with status
  • Redis Storage: Contributions stored for validation testing

Workflow Integration​

┌─────────────────────────────────────────────────────────────────┐
│                  Continuous Improvement Loop                     │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  task_audierne_docs (every 2h)                                  │
│        ↓                                                        │
│  Generate contributions from real docs                          │
│        ↓                                                        │
│  Create Opik dataset                                            │
│        ↓                                                        │
│  task_opik_evaluate (every 30min)                               │
│        ↓                                                        │
│  Evaluate Forseti accuracy                                      │
│        ↓                                                        │
│  Opik Optimizer                                                 │
│        ↓                                                        │
│  Improved prompts                                               │
│        ↓                                                        │
│  task_prompt_sync (midnight)                                    │
│        ↓                                                        │
│  Updated prompts in production                                  │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘