Setting up a knowledge base
Start from the documents, not from the settings.
Ask what they are
- Plain prose (policies, articles, manuals): the defaults are right.
- Scanned PDFs or images: they need OCR; pick a parser that reads images.
- Tables and spreadsheets (price lists, specs): larger chunks, so a row is not split from its header; a parser that keeps table structure.
- Short Q&A pairs (FAQ): small chunks, one question each.
- Code or logs: chunk on structure, not on characters.
The settings, in plain words
- Chunk size - how much text is one searchable piece. Bigger keeps context, smaller finds precise facts. Default first; change only with a reason.
- Top-k - how many pieces reach the agent per question. 4-8 is typical; more costs tokens and adds noise.
- Reranking - a second, smarter pass that reorders what was found. Turn it on when answers cite the wrong passage.
- Hybrid search - words and meaning together; helps with product codes and names.
Check it
After ingesting, search for three questions a real user would ask. If the right passage is not in the results, change one setting at a time.
