Travel planner agent recipe with granite 3.0 and LangChain Agents

[manojjahgirdar]

PR Checklist

Model Interaction

• [x] Flexible LLM platform support The platform should be easily switchable. Use LangChain or LlamaIndex.
• [x] Use prompt guide corresponding to the model For example for Granite 3.x Language Models

Data

• [x] Example data: Follow the example data guidance.

Notebook requirements

• [x] Notebook outputs cleared: Ensure all notebook outputs are cleared.
• [x] Automated testing: Add the recipe to the automated tests as described here
• [x] Test in Google Colab:
  • [x] Test that it works in Google Colab (Python 3.10.12).
  • [x] Colab has its own package set and Python version, so ensure compatibility.
• [x] Test locally:
  • [x] Ensure the code works in a fresh Python virtual environment (venv).
• [x] Standard access to secrets and variables Include !pip install git+https://github.com/ibm-granite-community/utils in the first code cell in order to make get_env_var available to accessing secrets and variables in the recipe.

Incoming References

• [x] README.md updates:
  • [x] Add a link to the recipe in the Table of Contents (ToC).
  • [x] Include a Colab button after that link.

GitHub

• [x] Commits signed: All commits must be GPG or SSH signed.
• [x] DCO Compliance: Developer Certificate of Origin (DCO) applies to the code, documentation, and any example data provided. Ensure commits are signed off.

[bjhargrave]

The ci build failed with an import error. I think you may be miss installing a needed package.

[manojjahgirdar]

Ok let me fix the import statements

[manojjahgirdar]

Hi @bjhargrave, I have fixed the import statement, how do I trigger the CI builds?

[manojjahgirdar]

It failed again, this time it says WATSONX_URL not found. If you see I have defined it correctly in the notebook.

[github-actions[bot]]

Testing Notebooks workflow launched on this PR: View run

[bjhargrave]

pull_request workflows do not have access to the secrets. I started a workflow which does.

[manojjahgirdar]

@bjhargrave can you rerun the CI pipeline now?

[github-actions[bot]]

Testing Notebooks workflow launched on this PR: View run

[manojjahgirdar]

@bjhargrave can you rerun the CI pipeline now?

[github-actions[bot]]

Testing Notebooks workflow launched on this PR: View run

[bjhargrave]

In looking at the code in the notebook, it does not seem to use the Granite 3 prompts for tool use: https://www.ibm.com/granite/docs/models/granite/#tool-usefunction-calling-weather-scenario

It just looks like vanilla langchain tool use.

[manojjahgirdar]

I just referred to the https://github.com/ibm-granite-community/granite-snack-cookbook/blob/main/recipes/AI-Agents/Agentic_RAG.ipynb recipe and tried to solve this new use case. I will try to change the prompt if that is absolutely required.

Changing the prompt may not give desired output, I can still give it a try.

[bjhargrave]

Understood. I am just trying to see if we should, in these Granite recipes, exploit the Granite prompt architecture for things like tool calling. @adampingel Do you have an opinion here? I know that many of the extant recipes were written before Granite 3 was released. But I would think that as we add new recipes, we should lean in to the Granite prompt architecture.

BTW, I have hacked on the existing Function_Calling snack to use the Granite prompt tool calling. I can make a PR for that update soon.

[manojjahgirdar]

Hi @bjhargrave, I have updated the prompt to use granite 3.0 prompt for tool calling. Can you rerun the pipeline to test it? Thanks.

[bjhargrave]

You also need to run the pre-commit hooks to clean up the notebook. There is also a readme conflict to resolve.

[github-actions[bot]]

Testing Notebooks workflow launched on this PR: View run

[manojjahgirdar]

Ran the pre-commit hook and resolved the readme conflict. Can you run the pipeline once again?

[github-actions[bot]]

Testing Notebooks workflow launched on this PR: View run