The Well of Kvasir: When the RAG Pipeline Learned to Listen
How a single email from a candidate revealed four distinct failures in four distinct layers — and why Separation of Concerns was the only way to fix them
Articles on Opik LLM observability and prompt optimization
Voir tous les tagsHow a single email from a candidate revealed four distinct failures in four distinct layers — and why Separation of Concerns was the only way to fix them
Three near-zero-cost interventions that improved our civic RAG pipeline more than any model change could
Comment le streaming, les threads et deux petits boutons ont transforme un outil de questions-reponses en un systeme qui apprend
Comment 60 captures d'ecran de reseaux sociaux sont devenues une memoire civique interrogeable en un apres-midi
"It looks impossible - but it's a hackathon. Cheers!"
This is the story of OCapistaine, a civic transparency AI built during the Encode "Commit to Change" Hackathon. It's a story of blocked pipelines, strategic pivots, 4,000 municipal PDFs, and the belief that AI can help citizens understand their local democracy.
Spoiler: We shipped it. Barely.

Sprint planning call between Johnny (@jnxmas) and Victor (@zcbtvag) to align on the Sunday midnight deadline. Key decision: pivot to Mistral Document AI + Batch + Agent for a rapid RAG prototype.
Quick catch-up between Johnny (@jnxmas) and Victor (@zcbtvag) covering Opik integration progress and OCR pipeline challenges.
Johnny showcased recent progress on Opik prompt optimization. A new architecture is in place where prompts are no longer hardcoded but managed via an Opik Prompt Library. The "Charter Validation" prompt has already been optimized using this system.
A new mock-up feature can automatically generate contributions (even with violations) from existing meeting reports. The goal is to create a robust dataset to test and improve the validation agent. However, this auto-generation currently produces repetitive content - a challenge that will need addressing by identifying and aggregating duplicate contributions.
Today we completed a major architectural milestone: modular prompt management for Forseti461. Each feature now has its own versioned prompt in Opik, enabling independent optimization and A/B testing.
From a single monolithic prompt to a clean separation of concerns — each Forseti feature can now evolve independently while sharing a common persona.
Forseti461 est un agent IA qui modère automatiquement les contributions citoyennes sur les plateformes de démocratie participative — approuvant uniquement les idées concrètes, constructives et localement pertinentes, tout en rejetant les attaques personnelles, le spam, les hors-sujets ou la désinformation, et en expliquant toujours ses décisions avec des retours respectueux et actionnables.
Ce week-end, Facebook nous a rappelé que la démocratie est fragile. Commentaires toxiques, attaques personnelles et diatribes hors-sujet ont envahi les discussions sur les enjeux locaux. Le signal se perd dans le bruit. Les citoyens se désengagent. Les voix constructives abandonnent.
Et si nous pouvions protéger le débat civique à grande échelle ?