TGTalker implementation

[shenyangHuang]

Summary / Description

running example, still have issues with backend.
There was some trouble with outlines API function change, have updated TGTalker to use the new one for structured json output

Related Issues: # (issue number)

Type of Change

• Bug fix
• New feature
• Breaking Change
• Refactoring
• Documentation update

Test Evidence

TO DO

• Unit tests
• Integration tests
• Performance tests

Questions / Discussion Points

Possible discussion to have it as submodule.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[shenyangHuang]

has LLM json output format issues such as the following on Qwen/Qwen3-1.7B. This is part of the model behavior and happens depend on the LLM

ERROR:__main__:Generation failed: Expecting ',' delimiter: line 1 column 37 (char 36)
{"destination_node": 135444228800000
ERROR:__main__:Generation failed: Expecting ',' delimiter: line 1 column 37 (char 36)
{"destination_node": 101298424033600
ERROR:__main__:Generation failed: Expecting ',' delimiter: line 1 column 37 (char 36)
{"destination_node": 481578640994762
ERROR:__main__:Generation failed: Expecting ',' delimiter: line 1 column 37 (char 36)
{"destination_node": 920892672325638

[shenyangHuang]

Currently missing the following:

• instruction set of the past 5 edges before prediction, also need to make user prompts of them and provide answer
• background set of the past 500 edges, this would need to be a separate dataloader, not entirely sure how it works directly in TGM. Basically it needs to update only after the standard batch size 200 dataloaders but instead I need access to the past 500 edges before the current batch of 200. The background set can be done as a custom hook

@ekmpa @Jacob-Chmura

[shenyangHuang]

@ekmpa can you provide a review for this PR as well