What problem does this solve?
I've been working in something similar from a few months now, and have a RAG with lance DB, and my breakthrough recently has been that the OpenAi 3072-dim embeddings makes everything work a 1000 times better.
In fact I'm open this issue because I expected MemPalace to beat one of my plugin skill for code review by a lot, but it did, it failed and the reason for it at least as identified by my agent is :
Mempalace missed it because ChromaDB's local embeddings (all-MiniLM-L6-v2, 384-dim) didn't score the query "TEXT primary key natural key surrogate integer" close enough to that entry. The smaller embedding model has less semantic resolution for database schema terminology.
Mempalace just does raw cosine similarity with no reranking — so lower-quality embeddings + no metadata boosting = missed finding.
What's the proposed solution?
Also anecdotic-ally I stopped using chroma I now use lanceDB and get better results.
Alternatives considered:
What problem does this solve?
I've been working in something similar from a few months now, and have a RAG with lance DB, and my breakthrough recently has been that the OpenAi 3072-dim embeddings makes everything work a 1000 times better.
In fact I'm open this issue because I expected MemPalace to beat one of my plugin skill for code review by a lot, but it did, it failed and the reason for it at least as identified by my agent is :
What's the proposed solution?
Also anecdotic-ally I stopped using chroma I now use lanceDB and get better results.
Alternatives considered: