Skip to main content

README

Current Manual Process for Handling Contributions​

The existing workflow is fully manual, relying on human oversight to ensure compliance and processing:

  1. Submission: Citizens anonymously submit ideas, proposals, or contributions via email to [email protected]. No direct web form exists yet, so email acts as the entry point.

  2. Manual Review (Garde-Fou Step): You personally check each incoming email against the project's charte (guidelines for neutrality, constructiveness, no toxicity, relevance to local issues, etc.). This involves reading the content, verifying compliance, and filtering out non-conforming submissions.

  3. Documentation of Compliant Contributions: For those that pass the charte check:

    • You copy-paste the content into a new GitHub issue in the repo audierne2026/participons.
    • You label or mark the issue as "conforme charte" to indicate it has been moderated.

  4. Creative Processing (Ocapistaine Step): Once marked compliant, you manually feed the contribution into the ocapistaine system (your emerging creative AI agent) for enrichment—e.g., generating suggestions, cross-referencing with local knowledge or external commune success stories, and producing an automated response or enhanced proposal.

This process ensures high-quality, guideline-aligned input but is time-intensive, scales poorly with volume, and depends entirely on your availability for review and processing.

Evolved Automated Process (Using the Discussed Stack)​

Leveraging the stack we've been building—FastAPI (backend/API), n8n (workflow orchestration and multi-agent logic), Opik (tracing, evaluations, and moderation metrics), hybrid LLM setup (local models for sensitive data + Mistral/cloud for advanced reasoning), and ocapistaine repo (crawling + creative agent)—we can automate most steps while preserving (and enhancing) privacy, moderation rigor, and creativity.

The goal: Shift to a semi- or fully automated pipeline with human-in-the-loop only for edge cases. Data privacy remains paramount—sensitive/personal contributions stay processed locally where possible, with triage before any cloud interaction.

Key Stack Components in the Evolution​

  • FastAPI: Exposes secure endpoints (e.g., for web form submissions or API triggers).
  • n8n: Orchestrates the full workflow (email polling → moderation → creative enrichment → output). Uses built-in AI Agent nodes for multi-agent setup.
  • Opik: Traces every step, runs automatic evaluations (e.g., moderation scores, bias detection), and logs for auditing.
  • Hybrid LLMs: Local (e.g., via Ollama/Mistral-7B) for garde-fou and sensitive data; cloud (Mistral Studio) for creative search/enrichment.
  • Additional Tools: Gmail integration in n8n for email ingestion; GitHub nodes for auto-creating issues; optional web form on audierne2026.fr pointing to FastAPI.

New Automated Workflow Steps​

  1. Submission (Improved Intake):

    • Short-term: n8n polls the Gmail inbox ([email protected]) periodically (e.g., every 5-10 minutes) using native Gmail nodes—fetches new emails anonymously.
    • Long-term: Add a simple web form on audierne2026.fr (built with FastAPI + basic frontend) for direct submissions. Anonymous by default (no login required), with optional email for follow-ups. Form posts to a FastAPI endpoint, triggering the n8n workflow via webhook.

  2. Automated Moderation (Garde-Fou Agent – First Line):

    • n8n routes the raw contribution (email body or form text) to the garde-fou agent (a dedicated AI node).
      • Uses a local LLM for privacy (sensitive local issues stay on-server).
      • Strict system prompt: Enforce the full charte (neutrality, no partisanship/toxicity, relevance, constructiveness).
      • Opik integration: @track decorator traces the moderation step; auto-runs evaluations (e.g., built-in ModerationMetric, custom bias/hallucination checks).
    • Outcomes:
      • Compliant: Proceed automatically; log as "conforme charte" in Opik.
      • Non-Compliant: Auto-reject with a polite templated email response (via n8n Gmail node) explaining why (e.g., "Does not align with charte guidelines on neutrality"). Flag for your review if score is borderline.
      • Human fallback: n8n notifies you (email/Slack) for manual override on ambiguous cases.

  3. Creative Enrichment & Response Generation (Ocapistaine Agent):

    • Only for compliant contributions: Route to the creative mind agent in n8n.
      • Hybrid: Local LLM handles core enrichment with private knowledge base (e.g., crawled local docs from ocapistaine).
      • Cloud LLM (Mistral) for external searches (success stories from other communes, best practices)—with secure triage (redact personal/sensitive info first).
      • Generates: Enriched proposal, automated response, cross-references, proactive suggestions.
      • Opik traces the full chain, scoring for relevance/creativity.

  4. Output & Documentation:

    • Auto-create a GitHub issue in audierne2026/participons (n8n GitHub node) with:
      • Original contribution.
      • Moderation status ("conforme charte").
      • AI-enriched version/response.
    • Optional: Post summary to the website (e.g., via FastAPI updating a database/feed) or send response email to submitter (if provided).
    • Proactive features (à la Jonas's inspiration): n8n scheduled jobs for daily briefs (emerging themes, win summaries).

Benefits of This Evolution​

  • Scalability: Handles higher volume without manual copy-paste.
  • Consistency: AI garde-fou applies charte uniformly; Opik provides auditable metrics.
  • Privacy: Hybrid routing ensures local-first processing.
  • Speed: Near-real-time responses encourage engagement.
  • Transparency: All traces in Opik dashboard for review; GitHub issues remain public record.

We're in early stages—start with n8n Gmail → moderation → GitHub proof-of-concept, then add web form and full hybrid. This aligns perfectly with civic tech goals for inclusive, safe dialogue ahead of 2026! If you'd like a sample n8n workflow export or FastAPI code snippet, just say the word.