I am wasting tokens because my agent guesses the wrong exact symbol names, a cheap embedding model like https://huggingface.co/minishlab/potion-multilingual-128M could be used to provide semantic search.
For reference, see this tool also from minishlab: https://github.com/MinishLab/semble
Ideally just copy the good parts of semble into codegraph for better cold-start query performance.
I am wasting tokens because my agent guesses the wrong exact symbol names, a cheap embedding model like https://huggingface.co/minishlab/potion-multilingual-128M could be used to provide semantic search.
For reference, see this tool also from minishlab: https://github.com/MinishLab/semble
Ideally just copy the good parts of semble into codegraph for better cold-start query performance.