tests/sync/test_run_sync.py

import json
import os
import numpy as np
import pytest
from common.db import init_db, get_paper
from common.vector_index import load_index
from sync.run_sync import run_sync

def test_run_sync_embeds_only_new_papers(tmp_path, monkeypatch):
    work_dir = str(tmp_path / "work")

    monkeypatch.setattr("sync.run_sync.download_snapshot", lambda repo_id, local_dir, token: {"papers": {}, "last_synced_at": None})
    uploaded = {}
    monkeypatch.setattr("sync.run_sync.upload_snapshot", lambda repo_id, local_dir, token: uploaded.setdefault("called", True))

    papers = [
        {"id": "p1", "title": "Title One", "abstract": "Abstract one", "authors": "A", "venue": "ACL", "year": 2023, "url": "http://x/p1", "bibtex": "@inproceedings{p1}", "pdf_url": "http://x/p1.pdf"},
    ]
    monkeypatch.setattr("sync.run_sync.iter_papers", lambda path: iter(papers))

    embed_calls = []

    class FakeClient:
        def __init__(self, *a, **kw):
            pass

        def embed_batch(self, texts):
            embed_calls.append(texts)
            return [np.ones(1024, dtype=np.float32) for _ in texts]

    monkeypatch.setattr("sync.run_sync.EmbeddingsClient", FakeClient)

    run_sync(
        anthology_path="/fake", hf_repo_id="org/repo", hf_token="tok",
        embedding_base_url="http://fake", embedding_api_key="key", work_dir=work_dir,
    )

    assert embed_calls == [["Title One\n\nAbstract one"]]
    assert uploaded.get("called") is True

    conn = init_db(os.path.join(work_dir, "papers.db"))
    paper = get_paper(conn, "p1")
    assert paper["title"] == "Title One"
    assert paper["bibtex"] == "@inproceedings{p1}"
    assert paper["pdf_url"] == "http://x/p1.pdf"
    index = load_index(os.path.join(work_dir, "index.faiss"))
    assert index.ntotal == 1
    assert paper["faiss_id"] == 0

    with open(os.path.join(work_dir, "state.json")) as f:
        state = json.load(f)
    assert "last_synced_at" in state
    assert state["last_synced_at"] is not None
    assert "papers" in state

def test_run_sync_raises_and_skips_upload_on_embedding_failure(tmp_path, monkeypatch):
    work_dir = str(tmp_path / "work")
    monkeypatch.setattr("sync.run_sync.download_snapshot", lambda repo_id, local_dir, token: {"papers": {}, "last_synced_at": None})
    uploaded = {}
    monkeypatch.setattr("sync.run_sync.upload_snapshot", lambda repo_id, local_dir, token: uploaded.setdefault("called", True))
    monkeypatch.setattr("sync.run_sync.iter_papers", lambda path: iter([
        {"id": "p1", "title": "T", "abstract": "", "authors": "", "venue": "ACL", "year": 2023, "url": ""},
    ]))

    class FailingClient:
        def __init__(self, *a, **kw):
            pass

        def embed_batch(self, texts):
            raise RuntimeError("embedding API down")

    monkeypatch.setattr("sync.run_sync.EmbeddingsClient", FailingClient)

    with pytest.raises(RuntimeError):
        run_sync(
            anthology_path="/fake", hf_repo_id="org/repo", hf_token="tok",
            embedding_base_url="http://fake", embedding_api_key="key", work_dir=work_dir,
        )
    assert "called" not in uploaded


def test_run_sync_twice_updates_faiss_id_mapping_on_reembed(tmp_path, monkeypatch):
    """Regression test for Fix #1: re-embedding an existing paper on a second
    sync run must supersede its old faiss_id rather than leaving a stale
    mapping, since add_vector always appends (never overwrites) in FAISS."""
    from common.db import get_paper_id_by_faiss_id

    work_dir = str(tmp_path / "work")

    # download_snapshot reads the on-disk state.json produced by the previous
    # run_sync call, mimicking real behavior across two syncs against the
    # same work_dir.
    def fake_download_snapshot(repo_id, local_dir, token):
        state_path = os.path.join(local_dir, "state.json")
        if os.path.exists(state_path):
            with open(state_path) as f:
                content = f.read().strip()
            return json.loads(content) if content else {"papers": {}, "last_synced_at": None}
        return {"papers": {}, "last_synced_at": None}

    monkeypatch.setattr("sync.run_sync.download_snapshot", fake_download_snapshot)
    monkeypatch.setattr("sync.run_sync.upload_snapshot", lambda repo_id, local_dir, token: None)

    class FakeClient:
        def __init__(self, *a, **kw):
            pass

        def embed_batch(self, texts):
            return [np.ones(1024, dtype=np.float32) for _ in texts]

    monkeypatch.setattr("sync.run_sync.EmbeddingsClient", FakeClient)

    papers_round1 = [
        {"id": "A", "title": "Paper A", "abstract": "", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/a"},
        {"id": "B", "title": "Paper B", "abstract": "original", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/b"},
    ]
    monkeypatch.setattr("sync.run_sync.iter_papers", lambda path: iter(papers_round1))

    run_sync(
        anthology_path="/fake", hf_repo_id="org/repo", hf_token="tok",
        embedding_base_url="http://fake", embedding_api_key="key", work_dir=work_dir,
    )

    conn = init_db(os.path.join(work_dir, "papers.db"))
    b_old_faiss_id = get_paper(conn, "B")["faiss_id"]
    assert get_paper_id_by_faiss_id(conn, b_old_faiss_id) == "B"

    # Second sync: B changes content (forcing re-embed/append), C is new.
    papers_round2 = [
        {"id": "A", "title": "Paper A", "abstract": "", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/a"},
        {"id": "B", "title": "Paper B", "abstract": "changed abstract", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/b"},
        {"id": "C", "title": "Paper C", "abstract": "", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/c"},
    ]
    monkeypatch.setattr("sync.run_sync.iter_papers", lambda path: iter(papers_round2))

    run_sync(
        anthology_path="/fake", hf_repo_id="org/repo", hf_token="tok",
        embedding_base_url="http://fake", embedding_api_key="key", work_dir=work_dir,
    )

    conn = init_db(os.path.join(work_dir, "papers.db"))
    b_new_faiss_id = get_paper(conn, "B")["faiss_id"]

    assert b_new_faiss_id > b_old_faiss_id
    assert get_paper_id_by_faiss_id(conn, b_old_faiss_id) is None
    assert get_paper_id_by_faiss_id(conn, b_new_faiss_id) == "B"


def test_run_sync_backfills_bibtex_without_reembedding(tmp_path, monkeypatch):
    """Bibtex and pdf_url must reach unchanged papers via the decoupled
    metadata write, not via re-embedding. Round 1 simulates pre-feature data
    (no bibtex/pdf_url keys -> stored as ''); round 2 supplies the same
    paper, now with both fields, but identical content hash -> unchanged ->
    zero embed calls, yet both fields end up populated. This is the
    guarantee that adding these fields does NOT trigger a full re-embed
    (which would orphan FAISS vectors, since add_vector only appends)."""
    work_dir = str(tmp_path / "work")

    def fake_download_snapshot(repo_id, local_dir, token):
        state_path = os.path.join(local_dir, "state.json")
        if os.path.exists(state_path):
            with open(state_path) as f:
                content = f.read().strip()
            return json.loads(content) if content else {"papers": {}, "last_synced_at": None}
        return {"papers": {}, "last_synced_at": None}

    monkeypatch.setattr("sync.run_sync.download_snapshot", fake_download_snapshot)
    monkeypatch.setattr("sync.run_sync.upload_snapshot", lambda repo_id, local_dir, token: None)

    embed_calls = []

    class FakeClient:
        def __init__(self, *a, **kw):
            pass

        def embed_batch(self, texts):
            embed_calls.append(list(texts))
            return [np.ones(1024, dtype=np.float32) for _ in texts]

    monkeypatch.setattr("sync.run_sync.EmbeddingsClient", FakeClient)

    # Round 1: legacy paper without a bibtex key (stored as '').
    papers_round1 = [
        {"id": "A", "title": "Paper A", "abstract": "", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/a"},
    ]
    monkeypatch.setattr("sync.run_sync.iter_papers", lambda path: iter(papers_round1))
    run_sync(
        anthology_path="/fake", hf_repo_id="org/repo", hf_token="tok",
        embedding_base_url="http://fake", embedding_api_key="key", work_dir=work_dir,
    )
    conn = init_db(os.path.join(work_dir, "papers.db"))
    assert get_paper(conn, "A")["bibtex"] == ""
    assert get_paper(conn, "A")["pdf_url"] == ""
    embed_calls.clear()

    # Round 2: same content (unchanged -> not re-embedded) but now with bibtex/pdf_url.
    papers_round2 = [
        {"id": "A", "title": "Paper A", "abstract": "", "authors": "", "venue": "ACL", "year": 2023, "url": "http://x/a", "bibtex": "@inproceedings{a}", "pdf_url": "http://x/a.pdf"},
    ]
    monkeypatch.setattr("sync.run_sync.iter_papers", lambda path: iter(papers_round2))
    run_sync(
        anthology_path="/fake", hf_repo_id="org/repo", hf_token="tok",
        embedding_base_url="http://fake", embedding_api_key="key", work_dir=work_dir,
    )

    assert embed_calls == []  # no re-embedding of an unchanged paper
    conn = init_db(os.path.join(work_dir, "papers.db"))
    paper = get_paper(conn, "A")
    assert paper["bibtex"] == "@inproceedings{a}"
    assert paper["pdf_url"] == "http://x/a.pdf"
    index = load_index(os.path.join(work_dir, "index.faiss"))
    assert index.ntotal == 1  # no orphaned second vector