ConferenceDB

Turn conference video recordings into a searchable local knowledge base.

22.5 hours of video → 2,726 searchable segments → answers in seconds

ConferenceDB takes raw conference recordings and builds a SQLite database with lexical search (FTS5), semantic search (sqlite-vec), and timestamped citations back to the original video. Every answer points to a moment you can verify.

Agent-native by design. The project ships an operational contract (AGENTS.md) that teaches any AI agent to run the full pipeline — from raw video to queryable knowledge base — with zero hand-holding.

What it looks like

# Semantic — find concepts across 22 hours of talks
$ .venv/bin/python src/query.py "how agents discover vulnerabilities" --mode vec

# Lexical — exact names, quoted phrases
$ .venv/bin/python src/query.py '"Trail of Bits"' --mode fts --limit 10

# Hybrid — FTS first, vector fallback
$ .venv/bin/python src/query.py "prompt injection defenses" --mode hybrid

Every result includes session, speaker, and timestamp — jump straight to that point in the video to hear it yourself.

Setup

python3 -m venv .venv
.venv/bin/pip install sqlite-vec sentence-transformers pytest

External tools (for processing new video):

Tool	Purpose
ffmpeg	Audio extraction from video
whisper-cli	Speech-to-text (whisper.cpp)

Pipeline

video files ──→ ffmpeg ──→ whisper-cli ──→ src/ingest.py ──→ SQLite
   .mp4          .wav         .srt           FTS5 + vec0      src/query.py

1. Gather metadata — conference, sessions, speakers → data/manifests/<slug>.json
2. Extract audio — ffmpeg to mono 16kHz WAV
3. Transcribe — whisper-cli to SRT, fix proper nouns
4. Ingest — src/ingest.py builds segments, FTS5 index, and 384d embeddings
5. Query — src/query.py or direct SQL/Python

Search modes

Mode	Flag	Best for
Lexical	--mode fts	Exact terms, speaker names, quoted phrases
Semantic	--mode vec	Conceptual queries, paraphrases
Hybrid	--mode hybrid	General use — FTS first, vec fallback

Data model

conferences → sessions → segments (20-40s timestamped chunks)
                ↓
             speakers (via session_speakers)

• segments_fts — FTS5 full-text index (BM25 ranking)
• segments_vec — sqlite-vec kNN index (all-MiniLM-L6-v2, 384 dimensions)

Full schema and query patterns: docs/semantic_layer.md

Tests

.venv/bin/python -m pytest tests/ -v

For AI agents

This project is designed to be operated by AI agents. Read AGENTS.md for the full operational contract.