Answers grounded in your company knowledge, with sources

Tables in PDFs become Markdown tables, scanned pages go through OCR, and images inside PDF and DOCX files are read by a model; spreadsheets are chunked by header plus row. Splitting method, max tokens and separator are adjustable, with a preview before import.

Choose mixed, semantic or keyword search; tune relevance from 0.1 to 0.95 and recall from 1 to 50; turn on a rerank model and knowledge-graph recall; and use retrieval testing to see which chunks come back.

Show sources as tags, a list or collapsed. The console Original Document panel highlights the passage in the PDF and jumps to the page; for spreadsheets it highlights the cell.

Google Drive is checked every 30 minutes and synced incrementally; websites are re-crawled on a schedule. Chunks can be added, edited, deleted, enabled or disabled, and are re-embedded on save.

Find a wrong answer in the logs, fix it with Train, and store it as a Q&A pair in the knowledge base you choose. It takes effect once embedding finishes.