What Knowledge Base does
Knowledge Base extracts text, preserves provenance, chunks documents and supports tenant-scoped retrieval. It is not model training. Extraction/indexing success and a useful retrieved answer are separate checks. Your selected agent still needs the appropriate knowledge/tool path.
Open Knowledge Base in the authenticated workspace. The deployed supported-type and limit metadata is the authority for your environment; administrators can inspect GET /api/v1/knowledge/supported-types with approved credentials.
Supported document families
| Family | Typical formats | Handling |
|---|---|---|
| Text/data | TXT, MD, CSV, TSV, JSON, YAML, XML, LOG | Native parsing/decoding |
| Web/email | HTML, HTM, EML | Body text extraction; active content removed |
| Native text first; OCR for low-text/scanned pages | ||
| Modern Office | DOCX, XLSX, PPTX | Paragraph, table, sheet and slide extraction |
| Legacy/OpenDocument | DOC, XLS, PPT, ODT, ODS, ODP | Isolated conversion, then extraction |
| Other | RTF | Text parsing |
| Images/scans | PNG, JPG/JPEG, TIF/TIFF, BMP, WEBP | OCR with confidence and page provenance |
Audio and video are not document uploads. File size, pages, decompressed pixels and archive expansion are bounded. Do not bypass limits by renaming an unsupported file.
Upload and verify
- Confirm data-owner approval, company context and the procedure's effective version.
- Choose a clearly named document, such as
Complaint Routing - reviewed 2026-09. - Upload and wait for a visible successful extraction/indexing outcome.
- Inspect extraction method, whether OCR ran and the confidence metadata.
- Search a known phrase and review source/page/sheet/slide details.
- Run a grounded question and verify it against the original document.
Each search result shows the document and, where recorded, its citation: page, sheet, cell range, paragraph or heading. Open excerpt shows the whole cited passage, with Previous and Next for the neighbouring passages. Where the deployment turns on query transformation (AGENTICORG_KNOWLEDGE_QUERY_TRANSFORM_ENABLED, off by default), How this was retrieved lists each retrieval step. Documents uploaded from this page are shared with the whole tenant. An API upload can name a domain (?domain=), and then users limited to other domains do not see that document.
If a duplicate dialog appears, choose Replace to index a new version before retiring the previous searchable copy, Keep both for a deliberate second copy, or Cancel to leave it unchanged. A failed replacement does not retire the original. Verify that search returns the intended version; do not leave conflicting policies current without an effective-date rule.
Replacement selects the current indexed revision ahead of failed or processing attempts, so retrying a failed upload does not leave the previous good copy current.
Domain-limited users cannot upload into another domain or replace/delete a document they cannot see. A hidden document is not disclosed by a duplicate-filename response.
| Status | Meaning | Next step |
|---|---|---|
| Processing | Stored, but not yet confirmed searchable | The page refreshes while processing. Do not repeatedly upload the same file. |
| Indexed | Indexing succeeded | Search a distinctive phrase and open its cited excerpt. |
| Failed | Indexing did not succeed | Inspect the error, then upload again and choose Replace. The prior good copy is retained. |
Where the administrator enables native upload recovery, temporary embedding timeouts, connection failures, rate limits and server errors receive at most four retries after the initial attempt, with delays of 30, 60, 120 and 240 seconds. The existing worker checks once a minute, so these are minimum delays, not completion guarantees. Invalid files, authorization errors and permanent indexing errors are not automatically retried. Large extractions outside the recovery storage bound and external RAGFlow indexing use the normal upload path instead. Your administrator can confirm which path is enabled.
- 1Approved upload
Size and type are validated before parsing.
- 2Extract or OCR
Native text is preferred; scanned pages are recognized where needed.
- 3Index with provenance
Chunks retain page, sheet, slide and freshness context.
- 4Search and review
Verify a known answer against the source before operational use.
Get better results from scans
Use clean, upright, complete scans with readable text. Avoid extreme compression, glare, truncated pages and mixed orientations. Keep the original for comparison. The OCR image includes India-first language packs, but operators must check installed languages and deployment settings for their material.
OCR confidence is diagnostic metadata, not the probability that a bank account number or monetary value is correct. Manually verify identifiers, totals, dates, negatives, handwritten fields and multi-column reading order. For a financial process, route ambiguous extraction to a reviewer; never reconstruct unreadable digits from context.
Errors and maintenance
Corrupt, oversized, unsupported or zero-text documents fail explicitly. A failed upload is not an empty valid policy. If OCR fails, ask the operator to check Poppler/Tesseract/LibreOffice availability and language packs. If extraction works but search fails, investigate embeddings and indexing separately.
Assign a content owner, review date and replacement/deletion procedure. Record which agents use the knowledge set. Use approved retention and access policies for customer documents; deleting a document and deleting derived evidence/history are different operations.