## Goal
For each question in the test set, submit:

1) a final answer (`answer`)  
2) the chunk IDs used as evidence (`cited_chunk_ids`)

Your score rewards **correct + grounded** answers with **precise citations**.

> This is NOT a tabular AUC/log-loss challenge.  
> Participants submit **answers + citations**, and the evaluation rewards **truthful grounding** and penalizes **hallucination behavior**.


Your score improves when:
- your answer matches the gold truth, AND
- your citations point to the correct supporting chunks, AND
- your answer content is consistent with what you cited.

## What you’ll learn (practical, job-relevant)
By participating, you’ll practice:

- Building a RAG pipeline (retrieve → answer → cite)
- Retrieval strategies (TF-IDF/BM25/embeddings)
- Citation selection policies (precision vs recall)
- Hallucination risk mitigation
- Benchmark-style evaluation thinking (groundedness metrics)

These skills map directly to real work in:
- LLM product engineering
- agentic workflows (Part II/III in the AgentEval series)
- search + retrieval systems
- evaluation and reliability of LLM outputs

---

## Data files
The competition provides four files:

- `corpus.csv`  
  A knowledge base split into text chunks.  
  Contains: `chunk_id`, `doc_id`, `title`, `doc_type`, `chunk_text`

- `train.csv`  
  Practice set including `answer` and `gold_chunk_ids`  
  (Use this to iterate and validate locally.)

- `test.csv`  
  Evaluation set (no answers)

- `sample_submission.csv`  
  Submission template with the correct format