The context
The deliberations, minutes and reports of a local authority represent several thousand pages produced each year. Officers lose considerable time finding a past decision, its context and its regulatory framework. Use of a consumer assistant was ruled out, both for confidentiality reasons concerning some in-progress reports and for political positioning on the digital sovereignty of the authority.
The need
An assistant that ingests the deliberative corpus, keeps it updated automatically at each new session, and lets both officers and elected representatives find a deliberation, an argument or a piece of administrative case law within seconds. Every answer had to point to the exact deliberation with its date and number. The infrastructure had to be hosted with a French cloud provider, with no dependency on a provider subject to the Cloud Act.
What we delivered
An automated ingestion pipeline for deliberations, with extraction, chunking and embedding computation. Hybrid storage combining pgvector for structured metadata and Qdrant for semantic search. Mistral model hosted on OVHcloud in the Gravelines region. FastAPI API exposing the assistant to the authority's internal applications. Architecture document setting out data residency, the absence of Cloud Act exposure, and the mapping to AI Act and GDPR obligations for a public service use. Training of internal officers included in the delivery.