Admin Settings → Documents configures the engines and defaults behind Document Parser, Extraction, Classification and Knowledge Bases. Changes affect how future documents are processed.
Content extraction
| Option | Effect |
|---|---|
| Content Extraction Engine | The primary engine used to read documents (built-in default, or an attached specialist engine — for example document-AI services or OCR-focused engines your deployment supports). |
| Engine connections | Specialist engines connect via their URL/key (for example a Document Intelligence endpoint and key). |
| Force OCR | Run OCR even on files with text layers. |
| Languages | OCR languages to improve accuracy. |
| Apply engine to modes | Choose which processing modes use a given engine. |
Upload and processing
| Option | Effect |
|---|---|
| Allowed File Extensions | Whitelist of uploadable types — restrict to what your workflows need. |
| Max Upload Count / Max Upload Size | Per-request upload limits; raise carefully — larger values consume processing capacity. |
| Keep Line Breaks | Preserve original line structure. |
| Extract Images / Disable Image Extraction | Pull embedded images out of documents. |
| Image Compression Width/Height | Downscale large images during processing (empty = no compression). |
Image description (multimodal knowledge)
| Option | Effect |
|---|---|
| Describe Pictures in Documents / Describe Images With VLM | Have a vision model write text descriptions of images, so image content becomes searchable in knowledge bases. |
| Image Description Source / Model Connection / Model | Which vision model performs the describing. |
| Image Description Prompt | Customise what the description focuses on. |
Retrieval defaults (RAG quality)
These values shape answer quality for every knowledge base:
| Option | Effect | Guidance |
|---|---|---|
| Chunk Size | Size of text pieces documents are split into. | Larger chunks = more context per hit but noisier; smaller = precise but can lose surrounding context. |
| Chunk Overlap | How much neighbouring chunks share. | Smooths splits. |
| Embedding Model / Engine | The model that converts text to vectors. | Changing it requires re-indexing all knowledge bases. |
| Embedding Batch Size | Processing batch for indexing. | Tune for speed on your cluster. |
| Top K | How many retrieved passages feed the answer. | More context, more tokens. |
| Top K Reranker / Score / BM25 Weight / Hybrid Search | Ranking refinements — combine keyword and semantic matching. | Hybrid search helps with exact terms like codes and part numbers. |
| Full Context Mode | Provide whole documents when they fit the context window. | Best for small document sets. |
| Bypass Embedding and Retrieval | Send content directly rather than retrieving. | Specialised use only. |
Maintenance actions
| Action | Effect |
|---|---|
| Reindex knowledge files | Re-process knowledge-base content — required after changing the embedding model. |
| Reset upload directory / Reset vector database | Administrative resets — destructive; export anything needed first. |
Change retrieval settings deliberately:
Chunking or embedding changes alter how existing knowledge bases answer. Re-test key question sets after any change, and re-index before rolling out.