How your agent knows things it was never trained on
A model doesn't know your return policy. RAG retrieves the right chunk and grounds the answer in it — with citations. Type a question and watch it happen.
A language model knows a staggering amount about the world and nothing about your business. It has never seen your return policy, your shipping zones, or that you take Yape. Ask it directly and it will do the most dangerous thing a support agent can do: guess, confidently.
The fix isn’t fine-tuning. It’s retrieval — pull the relevant piece of your knowledge into context, and make the model answer from that, not from memory.
Type a question below and watch the vector search re-rank the knowledge base on every keystroke, then ground an answer in the top chunks:
Type a question and watch the vector search re-rank live.
The answer is built only from what was retrieved — never invented.
Retrieval-augmented generation, in one breath
RAG has four moves:
- Chunk your documents into small, self-contained passages.
- Embed each chunk into a vector — a point in meaning-space.
- Search that space with the user’s question and take the top matches.
- Ground the answer: hand those chunks to the model and ask it to answer only from them, with citations.
The demo above uses simple keyword overlap so it can run in your browser — real retrieval uses embeddings and captures meaning, not just matching words. But the shape is exactly this: retrieve first, answer second.
”Grounded, not guessed” is the whole point
Notice what the answer panel does: it only ever says what a retrieved chunk supports, and it cites the source. Ask something the knowledge base can’t answer and it retrieves nothing — and a real agent, instead of inventing, asks a clarifying question or escalates.
That’s the difference between a demo and something you can put in front of customers. Citations aren’t decoration; they’re how a human can verify the agent, and how you keep it honest.
Retrieval quality is the product
Everyone can stuff text into a prompt. The hard, boring wins are in the retrieval: how you chunk, which embedding model you use, whether you re-rank the top results, how you handle freshness when a doc changes. Get that right and the same base model suddenly feels like it knows your company.
What Thaliq does with it
Drop your docs in. We chunk, embed with a portable vector store (pgvector), retrieve on every turn, and hand the model grounded context with citations — inside the same agentic loop as your tools and HITL. And because retrieval runs on your vector store, in self-host your knowledge never leaves your perimeter.
Your agent doesn’t need to be trained on your business. It just needs to be able to look it up.