RAG (retrieval-augmented generation) is how an assistant answers from your manuals, tickets, catalog, or API—not from a public chat model's memory. This page is the retrieval track inside AI implementation. Channel UX (Telegram, MAX, a site widget) lives on AI assistants and chat bots. If you only want a FAQ bot with no corpus, stop there. If the model must cite a price list that changes weekly, you are here.
I start from the source of truth: which files, which tables, who may see which slice. Then chunking rules (headings beat naive 500-token splits for contracts and 1C exports), embeddings, a vector store you can back up (often PostgreSQL + pgvector), and a retrieve-then-generate loop with citations. Low-confidence answers go to a human queue. Silent hallucination into accounting or support is a defect, not a “creative” reply.
A first vertical slice is usually one corpus and one audience: internal HR policy, a product FAQ, or one document type. We measure hit rate, citation correctness, and latency—not a vanity “it sounds smart” demo. Eval harness details sit on LLM gateway and eval. Agents that call tools (create a ticket, look up an order) are AI agents and MCP, not a second RAG page.
Stack: TypeScript/Node.js or Python for orchestration, mainstream embedding and chat APIs or on-prem when policy requires it. I do not fine-tune a 70B model as the default first milestone. Related engineering notes: Production RAG in 2026. Portfolio shape: thinklens-bot and cursor-telegram-integration show messengers plus retrieval-style grounding, not a slide-deck chatbot.
Out of scope unless named: scanning every shared drive “just in case,” training on customer data without a contract, or promising that RAG will replace your search team in a week. Timeline: a focused FAQ RAG often lands in 3–6 weeks after we have a clean document dump and access rules. Dirty PDFs, scans, and 1C metadata mapping add calendar time—OCR is document recognition.
What you get
- Inventory of sources, access roles, and “must not answer” topics
- Chunking, embeddings, retrieve/rerank, prompt with citations
- Staging corpus, eval set, and a short hypercare window after go-live
- Runbook: how to add a document, how to rebuild the index, how to read logs
