Model Config

Paperwise stores model connections and task routing per user. After signing in, open Settings > Model Config to configure the AI-backed parts of the product.

Supported connection types

Paperwise currently supports:

• OpenAI
• Gemini
• Custom (OpenAI-compatible)

Each connection can include:

• provider
• API key
• base URL when required
• default model

How task routing works

Paperwise lets you assign a connection and optional model override to each task:

• Metadata Extraction
• Grounded Q&A
• OCR

This means you can use a lighter or cheaper model for extraction and a stronger reasoning model for grounded Q&A.

Host-local providers with Docker

If Paperwise runs in Docker and LM Studio, Ollama, llama.cpp, or another OpenAI-compatible server runs on the Docker host, use host.docker.internal from Paperwise.

Use a base URL like:

http://host.docker.internal:1234/v1

Do not use localhost for this case. Inside Docker, localhost means the Paperwise container itself, not your host machine.

Both Paperwise services need access:

• api uses the provider for Test Connection and settings checks.
• worker uses the provider while processing documents.

On Linux Docker hosts, add this to your compose file if host.docker.internal is not already available:

services:
  api:
    extra_hosts:
      - "host.docker.internal:host-gateway"

  worker:
    extra_hosts:
      - "host.docker.internal:host-gateway"

If the model server runs on a different machine, bind that server to a network-reachable address and use that host name or IP instead.

Recommended first configuration

If you just want a working setup, start simple:

1. Add one provider connection.
2. Use that connection for Metadata Extraction.
3. Use that same connection for Grounded Q&A.
4. For OCR, choose either:
   • LLM if you want OCR handled through a multimodal model
   • Local Tesseract if you want OCR to stay local

OCR modes

LLM OCR

When OCR is set to LLM, Paperwise sends rendered page images to the selected model. This is usually better for scans, forms, image-heavy PDFs, and harder layouts.

Local Tesseract

When OCR is set to Local Tesseract, OCR runs locally using tesseract and pdftoppm. This is a good default for privacy-sensitive setups and clean printed scans.

Auto switch

Paperwise also supports an auto-switch mode so OCR is only used when direct text extraction looks weak.

Suggested starting models

These are practical starting points, not hard requirements:

Task	OpenAI example	Gemini example
OCR	gpt-5-mini	gemini-2.5-flash
Metadata extraction	gpt-5-mini	gemini-2.5-flash
Grounded Q&A	gpt-5.1	gemini-2.5-pro

If your documents are mostly clean text PDFs, start with the faster models and only move up when quality is not good enough.

User-verified self-hosted models

Paperwise users have reported successful local processing with this Ollama split:

Task	Ollama model
OCR	minicpm-v:latest
Metadata extraction	qwen3:8b
Grounded Q&A	qwen3:8b

Use a vision model such as minicpm-v:latest for OCR because Paperwise sends rendered page images for LLM OCR. Use a text model such as qwen3:8b for metadata extraction and grounded Q&A because those tasks work from extracted text.

Treat these as known-working starting points for self-hosted setups, not as the only supported local models. For custom providers, configure Ollama as an OpenAI-compatible connection, for example:

http://host.docker.internal:11434/v1

If Paperwise and Ollama run directly on the same machine outside Docker, http://127.0.0.1:11434/v1 is usually appropriate instead.

For text PDFs, keep Auto switch enabled when possible. It lets Paperwise try direct text extraction before OCR, which is usually faster than sending every page image to a multimodal model.

See Which models should I use? for more detailed starting recommendations and tradeoffs.

Common setup issues

• Upload blocked: configure Metadata Extraction first.
• Ask My Docs not available: configure Grounded Q&A first.
• OCR failures on scans: switch OCR to a stronger multimodal model or try Local Tesseract for cleaner documents.
• Custom provider not working: verify the base URL and API key in Model Config. For Docker plus host-local providers, use host.docker.internal instead of localhost.

Next: Q&A