oracle-ai-developer-hub
Soccer Analytics Agent Workshop
This directory contains the runnable 90-minute workshop materials for the World Cup 2026 soccer analytics agent.
OCI_GENAI_SETUP.md— OCI Generative AI setup notes.HACKATHON_GUIDE_ADB.md— original Reto Enseña / Autonomous Database guide retained for reference.
Final ML pipeline diagram
Editable draw.io source: ../docs/assets/ml_feature_pipeline.drawio
Detailed feature notes: ../docs/ML_FEATURE_PIPELINE.md
What the model uses
The prediction engine is not a toy prompt wrapper. The final predict_match path builds the same 92-feature row that trained the model:
| Feature family | Count | Why it matters |
|---|---|---|
| Original baseline: Elo, form/goals, H2H, match context | 40 | Long-term strength, recent trajectory, pair history, venue/rest/tournament setup. |
| Goalscorer intelligence | 12 | Scoring depth, star dependency, penalties, late goals, first-half scoring. |
| Momentum / psychology | 16 | Streaks, unbeaten runs, clean sheets, comebacks, draw tendency, blowouts. |
| Poisson expected goals | 8 | Independent goal-rate model with lambdas, win/draw probabilities, variance, overperformance. |
| Venue / geography | 5 | Altitude, confederation strength, intercontinental matchup effects. |
| Tournament context | 11 | World Cup form, competitive form, big-game factor, World Cup experience. |
| Total | 92 | Full XGBoost input row for live and cached predictions. |
Elo scoring note
The Elo system is replayed chronologically from the match dataset. Every team starts at 1500; non-neutral home teams receive a +100 expected-score bonus; and match importance is weighted by K-factor:
| Match category | K-factor |
|---|---|
| FIFA World Cup | 60 |
| Continental championship | 50 |
| Qualifier / Nations League / default competitive | 40 |
| Friendly | 20 |
Expected score:
E_home = 1 / (1 + 10^(-(R_home - R_away + HA) / 400))
Rating update:
R_new = R_old + K * G * (S_actual - S_expected)
G is the goal-difference multiplier: 1.0 for one-goal-or-draw outcomes, 1.5 for two goals, 1.75 for three goals, and 1.75 + (GD - 3) / 8 for four or more.
Retrieval and observability proof points
After predictions are created, scripts/load_langchain_vectors.py converts cached prediction rows and football facts into LangChain OracleVS documents in SOCCER_LANGCHAIN_DOCS using langchain-oracledb. The final Grok 4 chat should use hybrid_retrieve as the primary evidence path, then contrast it with semantic-only vector_search when requested.
Every /chat turn also writes ordered execution steps into Oracle through langgraph-oracledb OracleStore. Inspect them after a demo turn with:
curl http://localhost:8000/observability/<session_id> | uv run python -m json.tool
Expected events include turn_start, grounding_retrieved, model_response, tool_call, tool_result, and final_response.
Workshop demo prompt:
Use hybrid retrieval to explain the evidence for Spain vs Brazil, and contrast it with semantic-only memory.