How zhezh/kddcup2026_champion works
This repository is a benchmark data-agent application: it reads task directories, analyzes tabular, document, and video context, and produces task predictions. The documented entry point is zz_agent_v2.py; sampled code confirms its video-result and table-relevance stages and a solver fallback. The tree and excerpts support the main capability areas, but not every handoff among agents, so uncertain wiring is omitted.
Main entry points: zz_agent_v2.py Task runner · build_video_input.py Timeline assembly · video_result_agent.py Video answer review · doc_prepare.py Document structuring
Components 26
Task orchestration
- Task runner
src/data_agent_baseline/zz_agent_v2.py - Preflight checks
src/data_agent_baseline/preflight.py
Task understanding
- Table relevance
src/data_agent_baseline/agents_v2/table_relevance_agent.py - Query planning
src/data_agent_baseline/agents_v2/plan_agent.py - Document relevance
src/data_agent_baseline/agents_v2/doc_relevance_agent.py
Data preparation
- Document extraction
src/data_agent_baseline/agents_v2/doc_agent.py - Document structuring
src/data_agent_baseline/doc_tools/doc_prepare.py - Section structuring
src/data_agent_baseline/doc_tools/fanout_struct.py - Schema inference
src/data_agent_baseline/tools_v2/schema_inference.py - Record extraction
src/data_agent_baseline/tools_v2/doc_extractor.py - Structured document tables
Query and prediction
- Datasource setup
src/data_agent_baseline/tools_v2/scaffold_tool.py - Datasource runtime
src/data_agent_baseline/tools_v2/datasource_runtime.py - Data exploration
src/data_agent_baseline/tools_v2/explore_tool.py - Solver agent
src/data_agent_baseline/agents_v2/solver_agent.py - Solver execution
src/data_agent_baseline/tools_v2/run_solver_tool.py
Video analysis
- Video answer review
src/data_agent_baseline/agents_v2/video_result_agent.py - Timeline assembly
src/data_agent_baseline/video/build_video_input.py - Slide detection
src/data_agent_baseline/video/slide_coarse_det.py - Audio transcription
src/data_agent_baseline/video/audio_transcribe.py
Other components
- Benchmark operator
- Task files
- Task prediction
- Video and audio
- Whisper service
src/data_agent_baseline/asr/server.py - Model API
Connections 26
- Benchmark operator → Task runner: starts run
- Task files → Task runner: provides tasks
- Task runner → Preflight checks: checks tasks
- Task runner → Table relevance: requests analysis
- Task runner → Video answer review: reviews video
- Task runner → Solver execution: runs fallback
- Task runner → Task prediction: writes results
- Document structuring → Section structuring: structures documents
- Section structuring → Structured document tables: writes tables
- Document structuring → Structured document tables: stores tables
- Schema inference → Record extraction: supplies schema
- Datasource setup → Structured document tables: reads tables
- Datasource setup → Datasource runtime: builds datasource
- Data exploration → Datasource runtime: queries datasource
- Solver execution → Solver agent: invokes solver
- Solver agent → Task prediction: produces answer
- Video and audio → Timeline assembly: provides media
- Timeline assembly → Slide detection: extracts frames
- Timeline assembly → Audio transcription: transcribes audio
- Audio transcription → Whisper service: requests transcription
- Timeline assembly → Video answer review: provides timeline
- Table relevance → Model API: requests analysis
- Query planning → Model API: requests plan
- Document relevance → Model API: requests analysis
- Document extraction → Model API: extracts records
- Video answer review → Model API: reviews answer