Deep Research Alternative

This was started off of the reflex example but meant to be run locally.

Usage

To start, either run models locally or create ssh tunnel to remote Ollama models

ssh -L 11434:localhost:11434 -N ollamaserver

either set the env variables or use the included .env.sample:

❯ mise run show:secrets
[show:secrets] $ cat .env
JINA_API_KEY = "[redacted]"
SERPAPI_API_KEY = "[redacted]"
FIRECRAWL_API_KEY = "[redacted]"

in another terminal, start the reflex server:

uv run reflex run

Differences

• Use pydantic_ai for LLM interactions, allows multiple agents/providers which should be configurable through the app.
• Migrate to self-hosted alternative to jinaai and serpai, possible either a model or some alternative like:
  • https://github.com/intergalacticalvariable/reader
  • https://docs.firecrawl.dev/introduction

Other Deep Research Alternatives

• https://github.com/dzhng/deep-research
  • typescript and cli
• https://github.com/epuerta9/deep-research-py/
  • python and cli
• pydantic/pydantic-ai#901
• https://github.com/assafelovic/gpt-researcher
  • python, predates Deep Research
• https://huggingface.co/blog/open-deep-research

Notes / TODO

• If using Ollama, the models context length is likely to be the default for many model (2048). To extend the model to longer context which is necessary for various parts of this system to work. There is a simple helper script to extend the model in deepre/utils/ollama_helper.py and eventually that should be moved into a check done automatically that runs before the model is loaded.
• Using the result_type argument of run/agent doesnt seem to work when trying to cast the result for the models used. Should ideally be fixed but no obvious ways I've tried have worked.
• Should move the tool usage and delegation to agent delegation

Templates

• I am deciding if i should use toml for the templates or python files that i can define the result_type for the agent/model in as well.
• Toml is nice since simplifies