Local document AI
Run document extraction with local AI
Keep model inference on your computer without pretending the whole workspace is self-hosted.
The Dynamite Docs local companion sends document bytes from the browser to a service bound to loopback on your machine. That service prepares the document, calls your local Ollama model and returns structured results to the workspace for review.
Local inference is available on Hobby, Pro and Ultra. It is separate from BYOK, which connects an external provider account and requires Pro or Ultra. Account, workspace and token issuance remain hosted in both cases.
Set up local extraction. Local Ollama requires Hobby, Pro or Ultra, a running companion and a compatible local model.
What runs where
- Browser workspace: Starts the extraction and displays the editable result.
- Signed local companion: Accepts authenticated requests on loopback port 8756.
- Ollama: Runs the selected local model through its OpenAI-compatible API.
- Hosted account service: Issues short-lived tokens and keeps plan and workspace state.
Local models return the same review shape
The companion validates the model response before returning document type, schema, rows and confidence data to the workspace.
- Document type: invoice, statement, form or report
- Suggested schema: Named columns inferred from the file
- Rows: Structured records for review
- Metadata: Document-level context and fields
- Confidence: Result and field-level review signals
- Source content: Prepared document context for comparison
From file to reviewed result
- 1. Install Ollama and a model. Choose a model that supports the input type and fits the available memory and compute.
- 2. Start the companion. Run the bundled service on loopback. The default companion port is 8756 and Ollama defaults to 11434.
- 3. Connect from AI Settings. The hosted service issues a short-lived signed token and discovers models available through Ollama.
- 4. Extract and review locally. The browser sends the file to the companion, which prepares pages, calls Ollama and returns a validated result.
Select local models in the same AI panel
The workspace lists the chosen local model alongside hosted and own-key options. Local-only routing can prevent a document from falling back to an external provider.
The real routing rules panel used to restrict model selection.
The request stays on a short local path
The companion only listens on loopback by default. Every route requires a valid token, and browser origins are restricted to the official app, configured origins and local development.
| Step | Endpoint | Result |
| 1. Token | Hosted /api/local/token | Short-lived signed local token |
| 2. Models | Companion /v1/models | Ollama models available on the machine |
| 3. Extract | Companion /v1/extract | Validated schema, rows and confidence |
| 4. Review | Browser workspace | Editable extraction and export choices |
Local does not mean effortless or identical
Confidence
The workspace still shows confidence and editable results. Review money, identifiers and low-quality scans as you would with any provider.
Accuracy
Local model quality varies by model, quantization, document type and hardware. A smaller model may be faster but miss dense tables or small text.
Failure state
Extraction stops when Ollama is not running, the model is missing, the signed token expires, the companion cannot be reached or the model returns an invalid result.
Product boundary
The companion has a 150 MB upload limit, a default 120-second timeout and batches scanned PDF images within the model image limit. Hardware still sets practical speed and page limits.
Local and BYOK solve different problems
Use local inference when model calls should run on the machine. Use BYOK when an external provider account should be selected and billed directly.
| Option | Model runs | Plan access |
| Hosted | Dynamite Docs selected provider | Free, Hobby, Pro and Ultra |
| Local Ollama | User computer through the companion | Hobby, Pro and Ultra |
| BYOK | External provider account chosen by user | Pro and Ultra |
The processing boundary is specific
Document bytes used for local inference move from the browser to the loopback companion and then to Ollama on the same machine. The companion does not expose an unauthenticated public extraction endpoint.
Account, plan, workspace and token issuance still use the hosted Dynamite Docs service. Teams that need a fully self-hosted application should not treat the local companion as that product.
Questions about local document ai
Which plans include local Ollama?
Hobby, Pro and Ultra include the local companion. Free does not. This entitlement is separate from BYOK.
Does the local companion work offline?
Model inference runs locally, but the current product still uses the hosted account and workspace service for authentication, token issuance and account state.
Which Ollama model should I use?
Choose a model that supports the document modality and fits your hardware. Test it on representative files, then compare missing fields, table structure and correction time before using it for a batch.
Choose the next document task
Set up local extraction or compare processing, storage and integration limits.