Category Fingerprint Caching

[monapdx]

Add category fingerprint caching for efficient similarity scoring

Goal

Improve performance and scalability of similarity scoring by introducing cached category fingerprints, instead of comparing chats against every individual chat in a category.

---

Problem

Naive approach:

chat → every chat in every category

This becomes expensive as the dataset grows and will slow down the UI.

---

Solution

Introduce a category fingerprint:

A single vector representation of all chats in a category

Then compute:

chat → category fingerprint

This reduces comparisons from N chats per category → 1 vector per category

---

How It Works

• Each category maintains a cached vector (fingerprint)
• The fingerprint is built from:
  • Chat titles
  • User prompts (default)
• When a category changes, its fingerprint is recomputed

---

When to Recompute

Rebuild the fingerprint when:

• A chat is added to a category
• A chat is removed from a category
• A chat inside the category is edited (optional, can defer)

---

Storage

• Store fingerprints locally (SQLite or in-memory cache with persistence)
• Store metadata:
  • Last updated timestamp
  • Number of chats included
  • Vector size / method used (for future-proofing)

---

Implementation Notes

• Use the same vectorization method as similarity scoring (TF-IDF for MVP)
• Consider storing:
  • Raw vector
  • Or normalized vector (for faster cosine similarity)

---

Acceptance Criteria

• Each category has a cached fingerprint
• Fingerprints are only recomputed when necessary
• Similarity scoring uses fingerprints instead of per-chat comparisons
• No noticeable UI lag when rendering similarity scores
• Debug/logging option to inspect fingerprint rebuilds

---

Future Enhancements (NOT in this issue)

• Incremental updates instead of full recompute
• Support multiple fingerprint strategies (TF-IDF vs embeddings)
• Weighted fingerprints (recent chats matter more)

---

Why This Matters

• Keeps the app fast as data grows
• Enables real-time similarity scoring
• Makes the feature scalable without needing external APIs