RPI (Radial Priority Indexing) - A fork for self-correcting vector memory

[heydaytime]

I forked Qdrant to experiment with a feedback-driven retrieval policy called Radial Priority Indexing (RPI).

The problem: In repeated-query workloads (like NPC memory in games), bad context keeps surfacing. The model rejects it, you retry with a wider search, latency compounds.

The idea: Points live in trust shells (k=1 = high trust, k=N = low trust). Good retrievals get promoted inward, bad ones get demoted outward. Search starts at k=1 and falls through only if needed.

Key insight: Changing only a vector's magnitude doesn't change its cosine similarity direction - the semantic meaning stays stable. We use this to move points between Euclidean shell tiers for trust ranking while preserving their semantic relationships.

Results on 20 Newsgroups (5k docs, 5 epochs):

• Top-1 accuracy: 0.455 → 0.612 (+34%)
• Wider-scope retry rate: 0.54 → 0.42 (-23%)
• Avg candidate set size: 5000 → 3989 (-20%)

Not claiming this is production-ready or universally better - it's optimized for feedback-heavy, repeated-interaction workloads.
Repo: https://github.com/heydaytime/qdrant-rpi
Writeup: https://heydaytime.net/blog/rpi-semantic-cache-for-npc-memory
Would love any feedback from the community.

[generall]

Hm, if I understood your approach, you demote/promote results globally. So, if any query finds document A irrelevant, then is will be de-prioritized for all other queries as well, right?

[heydaytime]

Yeah, exactly. The design assumes a coherent query distribution, so global utility is meaningful. If a doc keeps getting passed over across this workload, it genuinely isn't useful to it. So we move it to further shells, and keep the closest k = 1 shell clean with good data.