An AI knowledge base to orient people, for health communities and local authoritiesThe right service, the right number, the right access conditions, in under thirty seconds.

Knowledge of a territory's offer, home-care services, respite schemes, drop-in sessions, community transport, prevention workshops, lives in a few people's memories, binders and outdated flyers. When they are away, orientation depends on who picks up. We built, for a professional health community, a tool that puts this knowledge within everyone's reach, with an AI assistant that helps but never decides.

Book a free diagnostic

Your situations

What front-desk and coordination teams experience

"My mother is losing her memory and I can't cope": the answer depends on who takes the call.

The respite-scheme flyer dates from 2023 and the number has changed.

A county-wide service that does not cover your area is suggested by mistake.

Three binders, twenty browser tabs, and nobody to keep it all current.

Newcomers need six months to know the territory.

The annual activity report is rebuilt from memory, for lack of records.

What we deliver

A graph of the territory

Services, organisations, venues, contacts, events, source documents, topics and audiences, linked on the real geography of your municipalities, with the source and date of every fact.

Plain-language search

One search bar: "physio near Braud", "caregiver respite". Results sorted by real distance, with phone, address, access conditions and a "check first" badge when the source ages.

A frugal AI assistant

It interprets vague requests and cites its sources; when the base does not know, it says so. No generated prose is ever handed to the people helped.

A shareable summary and pseudonymous statistics

Copy block, one-page PDF, prepared e-mail with the relevant documents. The activity log holds no personal data and feeds your activity report.

How it works

From your documents to a working tool

01
Collection
Your field documents, your reference websites, the professional directory. We extract, structure with AI assistance, and you validate.
02
Graph and geography
Municipalities, real adjacencies, territorial exclusions, topics: the base is built and verified by a battery of checks.
03
Testing with your teams
Panels of testers play first-day receptionists. We fix between each panel until a target usability score.
04
Go-live and living base
Hosting in France, training, then continuous enrichment by your partners: each organisation maintains and verifies its own information.

Why a graph, not a simple search engine

A search engine finds documents. A person at the front desk needs an answer: which service, how far, open when, under what conditions, and since when this information has been true. A graph links these facts together: the service, its organisation, its venue, the municipalities it covers, its exclusions, its source document and date. That is what makes it impossible to suggest a service that does not cover your municipality.

AI under source constraint

We did not plug a chatbot into PDFs. Interpreting a request is deterministic first: municipalities, ages, topics, synonyms from the real field vocabulary. The mid-sized AI only steps in when that is not enough, and every suggestion displayed comes from the graph, with its source and date. When the base does not cover a request, the tool says "no known organisation", it does not invent.

No data about the people helped

The tool describes the territory's offer, never patients or users. The activity log is pseudonymous by design: municipality, age group, topic, suggestions made. There is no field where an identity could be typed out of habit. Documents consulted are rendered client-side, without going through a third-party service.

What it costs

The prototype we delivered in Gironde runs for about €200 per month, hosting and AI included, for a whole territory of 62 municipalities. Building the base depends on the volume of documents; the production phase, hardening, backups, training, counts in days, not months.

This kind of project can be funded: in Nouvelle-Aquitaine, the Region supports the rollout of digital health solutions through calls for expressions of interest, and we drafted the corresponding application for our first client.

Transposable to your territory

There are more than 800 professional health communities in France, and as many territories where orientation knowledge is scattered. The architecture is generic: the knowledge schema is documented inside the base itself, geography initialises from public data, the ingestion pipeline replays on any corpus. A replication kit is planned with our first client.

The same tool serves a municipal social centre, a public-services hub, a local authority directing people to its social schemes.

References

French federation of health communities: the national framework.
The CNIL's guidance on AI: what is allowed with personal data, and how we avoid it.
The Model Context Protocol: the standard through which the assistant queries the graph.

Proof

What we delivered

For a professional health community in Gironde: a graph of 723 entities across 62 municipalities, 46 topics, an assistant that cites its sources, six tester panels, a usability score raised from 6.1 to 7.75 out of 10, and a regional funding application drafted.
A platform designed with no personal data, read-only access by default, encrypted secrets, tested backups.

Frequently asked questions

Questions we are asked

Our documents are PDFs and flyers, is that enough?

Yes, that is the usual starting point. We built the first base from 17 field documents and about twenty official websites. Extraction and structuring are AI-assisted, then validated by your referents before loading.

Who keeps the base up to date afterwards?

Your partners, each for their own information, with a freshness queue flagging what ages and automatic re-checking of websites. The goal is for maintenance to fit within existing coordination time, without a dedicated position.

Does the tool store health data?

No, by design. It describes the territory's offer, never people. Activity traces are pseudonymous, and there is no free field where an identity could be typed. A data-protection impact assessment can complete the setup before go-live.

Can it be opened to the public?

The tool is designed for professionals, internally, with authentication. Output to the public goes through the professional: a printout, a text message, an e-mail with the right documents. That is a choice: the people furthest from digital tools are those who most need well-equipped human guidance.

How long for a first territory?

A few weeks for a demonstrable prototype from your documents, then production rollout and supported experimentation over several months. We start with a demo on the existing base so you see the tool before deciding.

See the tool on a real territory

A thirty-minute demo on the existing base, then we look at what your documents would allow.