idref-alignment/README.md

# IdRef Alignment — Evaluation Script

`eval_idref_alignment.py` is a self-contained [uv](https://docs.astral.sh/uv/) script that
benchmarks the **Humatheque IdRef-Qualinka alignment API** against a golden dataset of
person → IdRef PPN mappings.

- **API under test:** [`idref-linker.smartbiblia.fr`](https://idref-linker.smartbiblia.fr/docs)
  (source: [gegedenice/humatheque-idref-qualinka-api](https://github.com/gegedenice/humatheque-idref-qualinka-api))
- **Evaluation dataset:** [`Geraldine/humatheque-vlm-sudoc-grounded-idref`](https://huggingface.co/datasets/Geraldine/humatheque-vlm-sudoc-grounded-idref)
- **Endpoint:** `POST /align/person`

---

## What it does

For every row of the dataset:

1. **Build the document context** from the `sudoc_record_templated` column:
   `title`, `subtitle`, `discipline`, `institution` (`granting_institution`),
   `doctoral_school`, `degree_type`, `year` (`defense_year`).
2. **Read the golden truth** from the `idref_persname_ppns` column — a list of
   `{ "Firstname Lastname": "PPN" }` pairs.
3. For **each person**, POST the name + context to `/align/person` and compare the
   returned PPN against the golden PPN.

The request body mirrors the API's `AlignPersonRequest` (see `build_align_payload`),
including the embedding model used for semantic similarity.

---

## Metrics

| Metric | Definition |
|---|---|
| **Accepted & correct** | `status == "accepted"` **and** `best_ppn == gold` — the production decision is right |
| **Accepted decisions** | rows where the API returned `status == "accepted"` (regardless of correctness) |
| **Top-1 correct** | `best_candidate.ppn == gold` — ranking quality, ignoring the accept threshold |
| **Candidate recall** | gold PPN appears anywhere in the returned `candidates[]` — recall of the candidate search |
| **Precision / Recall / F1** | of the "accepted" decision |
| **Status breakdown** | counts of `accepted` / `ambiguous` / `low_confidence` / `not_found` / `error` |

Each person also gets a `gold_rank` (1-based position of the gold PPN in the ranked
candidate list, or `null` if absent).

---

## Usage

Run directly from the Hub (no clone needed):

```bash
uv run https://huggingface.co/datasets/Geraldine/uv-scripts/resolve/main/idref-alignment/eval_idref_alignment.py
```

Or locally:

```bash
uv run eval_idref_alignment.py                       # full dataset, default settings
uv run eval_idref_alignment.py --limit 10 --concurrency 4
uv run eval_idref_alignment.py --embedding-model ""  # lexical similarity instead of embeddings
IDREF_API_KEY=xxx uv run eval_idref_alignment.py     # if the API requires an X-API-Key
```

### Key options

| Flag | Default | Description |
|---|---|---|
| `--api-base` | `https://idref-linker.smartbiblia.fr` | Base URL of the alignment API |
| `--dataset` | `Geraldine/humatheque-vlm-sudoc-grounded-idref` | HF dataset id |
| `--split` | `train` | Dataset split |
| `--embedding-model` | `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2` | Embedding model (empty string ⇒ lexical similarity) |
| `--max-candidates` | `20` | Max IdRef candidates per name |
| `--max-docs-per-role` | `20` | Max reference docs per role |
| `--reference-top-k` | `3` | Top-k references used as evidence |
| `--limit` | `0` (all) | Evaluate only the first N documents |
| `--concurrency` | `4` | Max concurrent API requests |
| `--request-timeout` | `90.0` | Per-request HTTP timeout (s) |
| `--out` | `idref_eval_results.jsonl` | Per-person JSONL output |
| `--summary-out` | `idref_eval_summary.json` | Aggregate summary JSON |

The API key and base URL can also be set via the `IDREF_API_KEY` / `IDREF_API_BASE`
environment variables. The deployed API is currently open (no key required).

---

## Outputs

- **`idref_eval_results.jsonl`** — one record per person with `name`, `gold_ppn`,
  `status`, `best_ppn`, `top1_ppn`, `top1_score`, `candidate_ppns`, `gold_rank`,
  and the correctness flags (`accepted_correct`, `top1_correct`, `topk_correct`).
- **`idref_eval_summary.json`** — aggregate metrics + status breakdown.
- A formatted summary table is also printed to the console.

---

## Notes

- The script follows the API's HTTP→HTTPS redirect automatically and retries failed
  requests with backoff.
- The dataset `thumbnail` (image) column is dropped on load to avoid an unnecessary
  Pillow dependency.
- Embedding-based requests are heavier server-side; for a quick smoke test use
  `--limit` and/or `--embedding-model ""`.