Option tables for the Document module. Administrator-configured defaults are set in Documents (Admin).
Parser options
| Option | Meaning | Notes |
|---|---|---|
| Extraction engine | The reader applied to files. | Structure-preserving engines for born-digital PDFs/Office; OCR engines for scans. |
| Force OCR | Apply OCR even when text exists. | Improves some scans; slower. |
| OCR languages | Languages to recognise. | Explicit languages beat auto-detect for non-English. |
| Keep line breaks | Preserve original layout lines. | Helps forms and tables. |
| Extract / disable image extraction | Pull embedded images out. | Enables image description downstream. |
| Image compression (W/H) | Downscale images during processing. | Saves space; leaves empty for no compression. |
| Allowed extensions | Uploadable file types. | Administrator whitelist. |
| Max upload count / size | Per-request limits. | Raise cautiously. |
File format support
| Format | Handling |
|---|---|
| PDF (text) | Highest fidelity, structure preserved. |
| PDF (scanned) | OCR applied. |
| Images (PNG/JPG…) | OCR for text; VLM description optional. |
| Office documents | Text and tables where present. |
| Text / Markdown | Light structuring. |
| Audio / video | Transcription paths where enabled. |
Extraction options
| Option | Meaning |
|---|---|
| Schema / field list | The named values to pull (e.g. invoice_number, total, supplier). |
| Preset schema | Ready-made field sets for common documents. |
| Confidence output | Per-field confidence, to route low-confidence cases to review. |
| Output format | Structured export (table/JSON/CSV). |
| Validation | Rules a value must satisfy (date format, numeric range). |
Classification options
| Option | Meaning |
|---|---|
| Labels / taxonomy | The categories to assign. |
| Label descriptions | Short definitions — the biggest accuracy lever. |
| Multi-label | Whether several categories may apply. |
| "Other" label | Catch-all to avoid forced wrong answers. |
Knowledge base / retrieval options
| Option | Meaning | Guidance |
|---|---|---|
| Chunk size | Size of indexed text pieces. | Balance context vs precision. |
| Chunk overlap | Shared text between chunks. | Smooths boundaries. |
| Embedding model | Converts text to vectors. | Changing it requires re-indexing. |
| Embedding batch size | Indexing throughput. | Tune for cluster load. |
| Top K | Passages retrieved per query. | More context, more tokens. |
| Top K reranker | Refined second-stage ranking. | Improves relevance. |
| BM25 weight / Hybrid search | Blend keyword + semantic matching. | Helps exact codes and names. |
| Score threshold | Minimum relevance to include. | Filters weak matches. |
| Full context mode | Feed whole documents when they fit. | Small sets only. |
| Image description (VLM) | Text descriptions of in-document images. | Makes image content searchable. |