Make your product data AI-readable — answers with a source, never a guess
An organisation with ~200,000 products behind clean APIs found that AI tools hallucinated, missed fields or built shadow copies of the database. We built an orchestration layer where AI never talks to the data directly natural language becomes a schema-validated query, and every answer is traceable to a source page.
Challenge
The problem pattern.
The data is not the problem. Organisations that manage large structured catalogues — hundreds of thousands of products, a hundred-plus data points each, open APIs — find that AI tools still fail on it. Generic copilots and app builders make too few calls, misread fields, hallucinate values, or quietly build their own copy of the database. For a system whose entire value is quality-assured data, wrong AI answers are unacceptable.
APIs are built for exact calls: article number in, fields out. AI users ask open questions — and that requires a layer of its own.
Solution
What we build.
An orchestration layer between users and data, built on one principle: the AI never talks directly to the data.
- Natural language is translated into a structured, schema-validated query that runs against an AI-optimised data access layer.
- Every answer is traceable to a concrete source page — ground truth, not a guess.
- The same architecture serves two channels: the organisation's own chat interface, and a partner API with per-customer keys so external parties can build their own integrations against the same endpoints.
The decisive design principle is schema-driven querying. The AI loads the dataset's schema at startup, knows exactly which fields, types and filters exist, and every generated query is validated against the schema before it runs. The AI cannot invent fields. Bad queries are caught before they reach the database.
Architecture
How it’s put together.
- Users & clients — browser sessions and external systems with their own API keys.
- Orchestration layer — a chat pipeline: normalisation → context reconstruction (follow-up questions rewritten as standalone) → intent detection (fast rules + LLM classifier) → routing to the right agent: greeting, explain, stats, compare/recommend, or structured query.
- Data access layer — FastAPI with a schema API and a responses API; columnar parquet storage chosen over RAG/vector databases for fast filtered reads of large structured datasets; nightly ETL refresh.
- The source data — untouched.
Key management follows the Stripe/OpenAI pattern: keys shown once, only hashes stored, per-integration keys so one partner can be paused without touching the rest.
Value
What it’s worth.
- For the data owner: a stronger position as the industry standard, partners building their own UIs on the infrastructure, and built-in verifiability — every AI answer links back to the source.
A demo on your real data within ~2 weeks is realistic. The architecture is a reusable template for any organisation with structured data and AI ambitions.
Questions about product data
Can an assistant answer from our product database without making things up?
Only if it is never allowed to answer from memory. The question is turned into a validated query against your own data, the assistant returns what the query returns and nothing else, and every answer carries the page it came from. Where the data holds no answer it says so rather than filling the gap. No second copy of the database is created along the way.