How vuongnguyen16/aic_2026 works
This is an end-to-end multimodal video retrieval platform. Its lifecycle runs from video keyframe extraction and filtering, through OCR and model-specific embeddings, into catalog/index construction; a user query is planned and routed across visual, OCR, and ASR evidence, then fused and reasoned over temporally before results reach the review UI or submission workflow. The graph preserves README capabilities whose implementation was not sampled, using only supported enclosing paths. Runtime wiring beyond the cited excerpts is based on the README’s explicit architecture; detailed internal calls are not inferred.
Main entry points: engine.py Search engine · app.py HTTP API · quality_filter.py Quality and dedup · fusion.py Rank fusion
Components 18
Video preparation
- Keyframe extraction
framme_extracting/src/framme_extracting/pipeline.py - Quality and dedup
framme_extracting/src/framme_extracting/quality_filter.py
Evidence enrichment
- OCR pipeline
baseline02/src/baseline02/ocr/pipeline.py - Embedding spaces
embedding - Catalog and indexes
Retrieval/rsys
Retrieval runtime
- Query planning
Retrieval/rsys/planning/planner.py - Query routing
Retrieval/rsys/planning/router.py - Search engine
Retrieval/rsys/engine.py - Visual OCR ASR branches
Retrieval/rsys/search.py - Rank fusion
Retrieval/rsys/fusion.py - Temporal reasoning
Retrieval/rsys/temporal.py - Optional reranking
Retrieval/rsys/rerank/pipeline.py
Review and delivery
- HTTP API
Retrieval/rsys/service/app.py - Evidence results
Retrieval/rsys/engine.py - Review workspace
Retrieval/web/src/App.tsx - Submission workflow
Retrieval/rsys/submission.py
Other components
- Video corpus
- Reviewer or agent
Connections 21
- Video corpus → Keyframe extraction: decodes frames
- Keyframe extraction → Quality and dedup: filters candidates
- Quality and dedup → OCR pipeline: provides keyframes
- Quality and dedup → Embedding spaces: provides keyframes
- OCR pipeline → Catalog and indexes: adds text evidence
- Embedding spaces → Catalog and indexes: builds vector indexes
- Catalog and indexes → Search engine: supplies indexed evidence
- Reviewer or agent → Review workspace: enters query
- Review workspace → HTTP API: sends requests
- HTTP API → Query planning: plans query
- Query planning → Query routing: routes query
- HTTP API → Search engine: executes search
- Search engine → Visual OCR ASR branches: searches modalities
- Visual OCR ASR branches → Rank fusion: provides ranked lists
- Search engine → Temporal reasoning: reasons over time
- Rank fusion → Temporal reasoning: provides ranked evidence
- Search engine → Optional reranking: offers candidate results
- Temporal reasoning → Evidence results: returns evidence
- Rank fusion → Evidence results: returns rankings
- HTTP API → Submission workflow: validates and exports
- Evidence results → Review workspace: displays evidence