[WIP] Hints retrieval

[ollmer]

1. First baseline - retrieve hints based on the relevance of the "semantic_keys" value to the given goal judged by llm. Choose 1 most relevant topic only.

Description by Korbit AI

What change is being made?

Add a new layered hint retrieval mechanism to the tool-use agent that employs a direct lookup or large language models (LLMs) based on a specified mode, update hints database and enhance logging and token usage handling.

Why are these changes being made?

These changes enhance the agent's functionality by allowing it to retrieve the most contextually appropriate hints, either directly or by leveraging an LLM, which is better suited for complex and nuanced filtering tasks. The updates to logging and exception handling in token usage improve robustness against unexpected API responses, ensuring smooth operation and debugging capability.

Is this description stale? Ask me to generate a new description by commenting /korbit-generate-pr-description

[korbit-ai]

Based on your review schedule, I'll hold off on reviewing this PR until it's marked as ready for review. If you'd like me to take a look now, comment /korbit-review.

Your admin can change your review schedule in the Korbit Console

[ollmer]

Baseline prompt example:

We're choosing hints to help solve the following task:
# Goal:

Create a filter for the list to extract all entries where:
    - "Assignment group" is "Software" and
    - "Caller" is "Fred Luddy".

You need to choose the most relevant hints topic from the following list:

Hint topics:
0. drop down menu
1. Filling up form
2. Search results
3. Dragging items
4. Shape and letters size comparison in miniwob
5. communicate answer in miniwob
6. Simbols of colors in miniwob
7. Reporting results in miniwob
8. Filling form in WorkArena
9. Filtering the list

Choose hint topic for the task and return only its number, e.g. 1. If you don't know the answer, return -1.

LLM response: 9