Suggest categories for uncategorized chats using similarity scores

[monapdx]

Suggest categories for uncategorized chats using similarity scores

Goal

Help users quickly scan uncategorized chats and identify likely topics by showing how similar each chat is to existing user-defined categories.

This is a discovery + triage tool, not a validation tool.

---

UI Behavior

For chats that are not yet assigned to a category, display 1–3 similarity indicators:

Likely: Coding (82%)
Also: GitHub (74%)

or more compact:

Coding 82% · GitHub 74%

Do NOT show this on already categorized chats (v1).

---

How It Works (Conceptually)

• User creates categories (e.g. Coding, Writing, Business)
• User assigns a minimum number of chats to each category (e.g. ≥5)
• Each category gets a text fingerprint/profile
• Uncategorized chats are compared against these category profiles
• The UI shows the strongest matches as suggestions

---

Important UX Constraints

• This is not auto-categorization
• This is not a ground truth label
• Language should reflect uncertainty:
  • “Likely”
  • “Similar to”
  • “Matches pattern”
• The feature should feel assistive, not authoritative

---

MVP Implementation Plan

Start simple and local-first:

• Use TF-IDF + cosine similarity
• Compare:
  • Chat title + user prompts (default)
  • (Optional later: full conversation)

---

Performance Strategy

Avoid expensive comparisons:

Instead of:

chat → every chat in every category

Do:

chat → category fingerprint

Where each category fingerprint is:

• A cached vector representation
• Built from all chats in that category
• Recomputed only when category assignments change

---

Acceptance Criteria

• Only appears when:
  • At least 1 category exists
  • Category has ≥5 assigned chats
• Shows top 1–3 category matches
• Scores are cached locally (SQLite)
• Updates when:
  • A chat is categorized/uncategorized
• Can be toggled on/off in settings
• Does not noticeably slow down UI

---

Future Enhancements (NOT in this issue)

• “Auto-suggest category” button
• Batch categorize similar chats
• “Similar chats” panel
• Embeddings-based similarity (optional)
• Adjustable weighting (title vs prompt vs full text)

---

Open Questions

• What minimum category size feels right? (3? 5? 10?)
• Should we normalize scores across categories?
• Should we hide scores below a threshold (e.g. <50%)?

---

Why This Matters

• Reduces manual sorting friction
• Makes large chat histories scannable
• Helps surface patterns the user didn’t explicitly define yet