While Databricks’ Deep Research product has yet to release, Langchain has released their Open Deep Research. In this article, we will explore how to leverage Databricks’ capabilities and run Open Deep Research using the foundation model APIs and built-in tracing to capture its research steps.
What is Open Deep Research?
First of all, the migration wouldn’t be possible without this free course from Langchain. I finished both the course and the migration in one evening. Thanks to both Langchain and Databricks made it so easy to accomplish Deep Research.
https://academy.langchain.com/courses/deep-research-with-langgraph
Many companies have developed Deep Research products and they are usually at the premium tier. For example, OpenAI has a Deep Research button, which we don’t get to use often as free users. Others like Gemini, Anthropic all have Deep Research available for a fee.
OpenAI Deep Research
The key differentiator of Open Deep Research is it is open source and the code is incredibly easy to understand. It doesn’t have a brain power, aka a Large Language Model, but it has all the logics required to perform deep research and that’s what make it so flexible because we can control not only the logic but the cost.
According to Langchain, Deep Research requires a lot of prompting tricks, as well as the “think tool”. All the prompts are also open source so we can tweak ourselves if we want. Quite literally nothing is held back with this premium product.
The architecture of Deep Research
Langchain has discussed this in quite detailed but in a nutshell, Deep Research has three phases:
- Scoping or clarification, similar to human asking clarifications before answering a question
- Research, there isn’t any mysteries about this as it is trying to leverage any tools available to do the web search until it reaches conclusion. However, Langchain also discussed that we need to handle the loop more carefully to control the time and cost so it will not go into an infinite loop of forever research.
- Write, finally we try to summarize or “compress”, all the findings into a report and the result is Deep Research.
Architecture of Deep Research
Onboarding to Databricks
The only major change is probably migrate from init_model to ChatDatabricks and of course use MLflow’s autolog(). But because Langchain wants us to run on their platform, the instruction of how to run locally wasn’t provided, that was another tricky thing.
To accelerate access to Deep Research on Databricks, below I have provided the code to run Open Deep Research (regardless if you use ChatDatabricks):
Enjoy!
AUTHOR - FOLLOW
Jason Yip
Director of Data and AI, Tredence Inc.
