anibon-search
Installation
SKILL.md
ANIBON Search Skill
1. Overview & Principles
Purpose
This skill retrieves specific moments from the ANIBON livestream archive — a corpus of timestamped markdown files and compressed transcripts covering Boat's (ปู่โบ๊ต) livestreams. It answers questions like "Which stream did Boat talk about X?" or "Find the clip where Boat criticized Y."
Hybrid Retrieval Architecture
The search pipeline is deterministic and uses no LLM at retrieval time:
| Stage | Mechanism |
|---|---|
| Lexical / Fuzzy | SQLite FTS5 trigram index over tokenised terms (Thai + Latin). Handles partial matches, common misspellings, and CJK/Thai substring queries. |
| Semantic | Pre-computed vector embeddings (sentence-level) compared via cosine similarity against the query's semantic field. |
| Fusion | Reciprocal Rank Fusion (RRF) combines lexical and semantic rank lists into a single ordered result set. |
| Moment Collapse | Results are collapsed to 600-second windows so that multiple hits within the same 10-minute segment return as one moment entry rather than duplicate rows. |