SKILL: Find (Discovery & Retrieval)

Setup for Find Module

1. Main Goal

Maintain and query a high-quality library of tactical round documents. Users query this pre-indexed dataset (e.g., VCT Champions 2025) to retrieve precise, timestamped VOD results.

2. Capabilities

• •Recursive Crawling: Drill down into rib.gg events using Playwright to find match IDs.
• •Hybrid Retrieval: Extract intent (Team, Map, Round Type) using Gemini and apply hard metadata filters before semantic search.
• •Data Fidelity: Track running match scores and tactical round ceremonies (Thrifty, Flawless).
• •Interactive VOD: Auto-seek YouTube players based on game-time offsets.

3. Data Schemas

• •Input: Natural Language Query (e.g., "PRX flawless rounds on Sunset").
• •Intent: { team_slug: 'paperrex', map: 'Sunset', round_type: 'flawless' }.
• •Output: MatchResult[] with vod_timestamp, video_id, team_a, team_b, and score_a/b.

4. Operational Constraints

• •Deduplication: Always deduplicate results by round_id from the source.
• •Conditional Threshold:
  • •Filtered queries (team/map/round_type present): No threshold, metadata acts as hard pre-filter
  • •Semantic queries (no filters): Apply cosine similarity threshold of 0.5
• •VOD Precision: Timestamps must be calculated via vod_start + (round_gametime - match_start_gametime).

5. Technical Stack

• •Database: Supabase PostgreSQL with pgvector extension (ChromaDB fallback)
• •Embeddings: Gemini text-embedding-004 (768 dimensions)
• •Search Function: match_rounds RPC with conditional threshold logic
• •Intent Detection: Gemini 3 Flash for query parsing