Experiment: long-term memory recall

luminai-companion / filament

A Python API for talking with KoboldAI with request and response data manipulation.

GNU General Public License v3.0

3 stars 1 forks source link

Experiment: long-term memory recall #7

Closed Pleasant-Cry closed 1 year ago

Pleasant-Cry commented 1 year ago

This idea is a bit vague, but we could implement an experiment where:

Each statement in a bank of "long-term memories" is embedded using something like WordNet or Word2Vec and stored
Statements in the prompt are embedded using the same approach
A metric like cosine similarity is applied to determine semantic similarity between prompt statements and memories
Top K memories exceeding some similarity threshold are injected into the prompt before sending to the LLM

garrus0xff commented 1 year ago

Or we can use sentence-transformers - https://www.sbert.net/examples/applications/semantic-search/README.html - ideally trying various approaches.

Pleasant-Cry commented 1 year ago

Or we can use sentence-transformers - https://www.sbert.net/examples/applications/semantic-search/README.html - ideally trying various approaches.

Absolutely trying sentence-transformers right now -- thanks for putting that idea in my head. I posted some quick experimental results in the model middleware post channel on discord, if you're still around there. We miss you!