We want to make our own knowledge searchable for AI but do not know which parts we need

A working RAG (retrieval-augmented generation) pipeline has four blocks: sources, ingest with chunking, vector database, and the query layer that feeds the model. For SMB keep each block as simple as possible and only extend when you hit a real bottleneck.

Try this first

1. 1Sources: decide which documents really matter. Start with SharePoint, Drive or one specific folder with handbooks and procedures. Not everything at once, or you can never debug what is in there.
2. 2Ingest: pick a tool or script that fetches sources, converts to text, splits into chunks, and stores them as embeddings. For SMB an open-source ingest or a SaaS like Vectorize or Carbon is usually enough.
3. 3Vector database: from a few thousand to a million chunks, pgvector on Postgres, Qdrant and Pinecone are roughly interchangeable. Pick the one closest to where your other data already lives.
4. 4Query layer: a thin app or n8n flow takes the user question, fetches top-k chunks, pastes them as context, and sends to the model. Show sources under the answer, not just the answer.
5. 5Eval: test with twenty real questions where you know the right answer. Only then do you know the pipeline works. Add questions with the right answer being 'not in our docs' too, hallucination testing matters.

None of the above fits?

Describe your situation below. We pass your input plus the steps you already saw to our AI and return tailored next-step advice. If it's too risky to DIY, we'll say so.