Guided Intake Chat
A conversational procurement intake for a European bank, built as a prototype. A colleague describes what they need in
plain words, and the chat:
- finds the buying route with fixed rules (catalogue, contract call-off, low-value purchase, or new demand with three
quotes or an RFP);
- asks only what that route needs, one question at a time;
- lets the requester choose the Level 4 category from 2–3 candidates, then classifies the request (ICT service under
DORA, outsourcing under EBA/GL/2019/02);
- hands procurement a service description, or a short purchase request, plus a record of everything captured.
The AI holds the conversation. It never decides the route or the classification.
Live at https://guided-intake-chat.vercel.app (access code required). Dummy data only: the AI runs on free provider
tiers.
Documentation
| Document | Contents |
|---|
| docs/PRD.md | Problem, users, scope, routes and thresholds, requirements, success measures, known gaps, next steps |
| docs/ARCHITECTURE.md | Components, the turn flow, AI calls and guardrails, classification, reference data, server, deployment |
| docs/DESIGN.md | Principles, layout, components, visual language, voice, states, accessibility |
| How it works (link in the app) | The full specification generated from the code: every rule, question and exact prompt |
| Decision tree (link in the app) | The same rules for business readers: the whole tree as one picture, the yes/no ladders for the route and the checks, which question drives which decision, and worked examples. Generated from the code; downloads as Markdown (with a Mermaid diagram) or as an SVG picture |
What is in this repository
| Path | Role |
|---|
index.html | The whole application, with no build step. It includes the reference data (one JSON block), the How it works specification and the Decision tree |
api/llm.js | Vercel serverless function. It holds the AI keys and the access code, and tries Gemini, Groq, Mistral and OpenRouter in turn |
vercel.json | Function timeout (60 s) and security headers |
docs/ | Product, architecture and design documentation |
tests/, scripts/, package.json, playwright.config.js | API and end-to-end tests, the local server, the mock providers and the data import script |
.github/workflows/ci.yml | CI: runs all tests with mock providers on every push and pull request |
The test data the taxonomy was imported from and task notes stay on the developer's machine
(see Architecture §10).
Develop and test
npm ci
npx playwright install chromium
npm test # API tests + Playwright, mock providers, no keys or network
scripts/verify.sh # runs the CI checks locally; runs automatically on `git push`
npm start # local server on http://localhost:3000 (npm run start:mock for fake providers)
Deploy
Set up Vercel:
- Push to
main: Vercel deploys to production.
- Framework preset: "Other". No build command.
Environment variables:
- at least one of
GEMINI_API_KEY, GROQ_API_KEY, MISTRAL_API_KEY, OPENROUTER_API_KEY;
ACCESS_CODE;- optional
LLM_PROVIDER (which provider goes first) and GEMINI_MODEL, GROQ_MODEL, MISTRAL_MODEL,
OPENROUTER_MODEL;
- optional
RATE_LIMIT_PER_MIN (default 30 AI calls per IP and minute) and CHECK_LIMIT_PER_MIN (default 5).
Check providers: /api/llm?check=1 with the access code in the x-access-code header.
Settings
The thresholds are in CONFIG.thresholds at the top of the script in index.html. They are set by procurement and
cannot be changed by requesters:
| Setting | Value |
|---|
| Low-value limit | €5,000 |
| Business-led sourcing limit | €50,000 |
| Material spend review | €1,000,000 |
| Quality minimum | 75 |
| Level 4 candidates shown from | 75% match (2–3 shown, the requester chooses) |
