Adding magics functionality to the litellm implementation
#1437
Conversation
There was a problem hiding this comment.
@srdas This is a great start, thank you! Just one extra comment below.
In addition, I've opened some follow-up issues (titled [magics] ...) for you to explore. They can be addressed in separate PRs.
This is already a fantastic start, and I'm excited to merge this once the last comment is addressed.
| self.transcript.append({"role": "user", "content": prompt}) | ||
| self.transcript.append({"role": "assistant", "content": output}) |
There was a problem hiding this comment.
Can you bound this list to be of length self.max_history * 2? Right now, the ai() method is manually truncating self.transcript every time it appends it to the chat history. It would be more reliable to add some logic to bound the length of self.transcript in the _append_exchange() method itself.
There was a problem hiding this comment.
Yes, this is now done. Tested everything to make sure it is working. Here is a notebook if you want to test all the magics features (unzip to use).
magics_litellm.ipynb.zip
Refactored various functionality for the
%aiand%%aimagics to use the newlitellmbackbone added in PR #1426 . Fixes Issue #1432run_ai_cellto invokelitellminstead oflangchain.registercommand toaliaswhich is more natural when registering an alias for amodel_id.deletecommand todealiaswhich is more natural.updatecommand as it is working well and is convenient.helpis working and shows the updated commands.listcommand to show model IDs and aliases.resetcommand is working.default_language_modeltoinitial_language_model.aliasestoinitial_aliases.errorcommand tofix, updated it's handling..envapproach.See the demo below showing all of it working:
Screen.Recording.2025-07-30.at.2.30.01.PM.mov
Nice to have (among other items):
formatoptions to be much more stable using agentic capabilities with better prompts.widgetsas another output format option.TODOcommented items inmagics.pyfor consideration.