Reach and enforce 100% coverage

Cover the remaining paths in the client, context, downloads, title
generation, document tools and store models, and add fail_under=100 so
uncovered lines fail CI.

Six lines that no test can reach get a pragma with its reason: the docling
import guard, the nameless PDF attachment, the FS symlink OSError guard that
resolve(strict=False) absorbs, the docling bbox and LanceDB document-id
shape guards, the tag-retention branch vacuum makes unreachable, and Monty's
Rust-thread print callback.

Fix test_find_config_file_user_config, which wrote its config into the cwd it
had chdir'd to, so the cwd branch answered first and the user-directory
lookup it names was never exercised.

CHANGELOG.md

@@ -1,6 +1,10 @@
 # Changelog
 ## [Unreleased]
 
+### Changed
+
+- Test coverage is enforced at 100% via `fail_under` in `[tool.coverage.report]`.
+
 ### Fixed
 
 - `Store.set_haiku_version` stamps the store's own config into a recreated settings row instead of the process-global `Config`.

haiku_rag_slim/haiku/rag/client/documents.py

@@ -570,7 +570,8 @@ def _extract_pdf_attachments(
             for i in range(attachment_count):
                 att = pdf.get_attachment(i)
                 name = att.get_name()
-                if not name:
+                if not name:  # pragma: no cover - pypdfium2 cannot produce a
+                    # nameless attachment, so no craftable PDF reaches this.
                     continue
                 data = bytes(att.get_data())
                 child_uri = f"{parent_uri}#attachment={quote(name, safe='')}"

haiku_rag_slim/haiku/rag/client/downloads.py

@@ -38,7 +38,8 @@ async def download_models(
         yield DownloadProgress(model="docling", status="start")
         await asyncio.to_thread(download_models)
         yield DownloadProgress(model="docling", status="done")
-    except ImportError:
+    except ImportError:  # pragma: no cover - docling is installed in the test
+        # environment, and get_package_versions depends on it.
         pass
 
     # HuggingFace tokenizer

haiku_rag_slim/haiku/rag/store/engine.py

@@ -606,7 +606,9 @@ class Store:
             for tag in tags.values()
             if tag["version"] in timestamps
         ]
-        if not tagged:
+        if not tagged:  # pragma: no cover - a tag's version is never absent from
+            # list_versions: vacuum retains every version at or after the
+            # oldest tag, so a tagged version is never cleaned away.
             return retention
 
         # LanceDB version timestamps are naive datetimes in local time.

haiku_rag_slim/haiku/rag/store/models/chunk.py

@@ -74,7 +74,8 @@ class ChunkMetadata(BaseModel):
                 continue
             for prov_item in prov:
                 bbox = getattr(prov_item, "bbox", None)
-                if bbox is None:
+                if bbox is None:  # pragma: no cover - docling's ProvenanceItem
+                    # always carries a bbox; guards against a shape change.
                     continue
                 bounding_boxes.append(
                     BoundingBox(

pyproject.toml

@@ -146,6 +146,7 @@ omit = [
 
 [tool.coverage.report]
 show_missing = true
+fail_under = 100
 exclude_also = [
     "if TYPE_CHECKING:",
     "@abstractmethod",

tests/store/test_document_items.py

@@ -1003,3 +1003,35 @@ async def test_replace_for_document_with_no_items_deletes_existing(temp_db_path)
         await repo.replace_for_document("doc-1", [])
 
         assert await repo.get_all_items("doc-1") == []
+
+
+class TestExtractItemTextFallbacks:
+    def test_table_returns_none_when_serialization_fails(self):
+        """A serializer that raises leaves the table with no extractable text
+        rather than aborting the extraction pass."""
+        doc = _doc_with_tables(1)
+
+        class _Boom:
+            def serialize(self, item):
+                raise RuntimeError("serializer exploded")
+
+        assert extract_item_text(doc.tables[0], doc, get_serializer=_Boom) is None
+
+    def test_file_backed_picture_has_no_inline_bytes(self):
+        """A picture whose ImageRef points at a file rather than a data: URI
+        carries nothing to decode."""
+        from docling_core.types.doc.document import ImageRef
+
+        from haiku.rag.store.models.document_item import _decode_picture_bytes
+
+        doc, pic = _doc_with_captioned_picture("caption")
+        pic.image = ImageRef.model_validate(
+            {
+                "mimetype": "image/png",
+                "dpi": 72,
+                "size": {"width": 1, "height": 1},
+                "uri": "file:///tmp/picture.png",
+            }
+        )
+
+        assert _decode_picture_bytes(pic) is None

tests/store/test_restore.py

@@ -112,8 +112,10 @@ async def test_restore_safety_tag_name_collision(temp_db_path, monkeypatch):
         await store.create_tag("release-1")
         await store.create_tag("before-restore-20260715T143012Z")
 
+        await store.create_tag("before-restore-20260715T143012Z-2")
+
         safety_tag = await store.restore_tag("release-1")
-        assert safety_tag == "before-restore-20260715T143012Z-2"
+        assert safety_tag == "before-restore-20260715T143012Z-3"
 
 
 @pytest.mark.asyncio
@@ -406,3 +408,24 @@ async def test_wait_protected_returns_result_on_same_tick_cancellation():
     result, cancelled = await outer
     assert result == "done"
     assert cancelled is True
+
+
+@pytest.mark.asyncio
+async def test_wait_protected_reraises_when_recovery_itself_is_cancelled():
+    """If the recovery coroutine ends cancelled there is nothing to wait for,
+    so the cancellation propagates instead of looping forever."""
+    import asyncio
+
+    from haiku.rag.store.engine import _wait_protected
+
+    async def self_cancelling_recovery() -> str:
+        current = asyncio.current_task()
+        assert current is not None
+        current.cancel()
+        await asyncio.sleep(0)
+        return "unreachable"
+
+    outer = asyncio.create_task(_wait_protected(self_cancelling_recovery()))
+
+    with pytest.raises(asyncio.CancelledError):
+        await outer

tests/store/test_v0_50_0_migration.py

@@ -126,6 +126,19 @@ class TestV0_50_0Migration:
                             }
                         ),
                     ),
+                    # Matches the LIKE on the quoted-key form, but only nested —
+                    # there is no top-level key to rename.
+                    LegacyDocumentRecord(
+                        id="nested-only",
+                        content="x",
+                        uri="u3",
+                        metadata=json.dumps(
+                            {
+                                "raw_headers": {"etag": "abc"},
+                                "source_revision": "v3",
+                            }
+                        ),
+                    ),
                 ],
             )
 
@@ -141,6 +154,10 @@ class TestV0_50_0Migration:
                 "my_etag_key": "v",
                 "source_revision": "v2",
             }
+            assert by_id["nested-only"] == {
+                "raw_headers": {"etag": "abc"},
+                "source_revision": "v3",
+            }
 
     async def test_unparseable_metadata_skipped_without_crashing(self, temp_db_path):
         """A row with malformed JSON in `metadata` must not abort the whole

tests/test_chunk.py

@@ -482,6 +482,9 @@ async def test_chunk_repository_get_by_id_and_list_all_pagination(temp_db_path):
         assert len(first) == 1
         assert first[0].id == everything[0].id
 
+        # offset is applied even when it selects the whole set
+        assert len(await client.chunk_repository.list_all(offset=0)) == len(everything)
+
         if len(everything) > 1:
             second = await client.chunk_repository.list_all(limit=1, offset=1)
             assert len(second) == 1
@@ -526,4 +529,4 @@ async def test_process_search_results_rejects_unknown_score_column(temp_db_path)
                 return pd.DataFrame([{"id": "c1", "content": "x", "metadata": "{}"}])
 
         with pytest.raises(ValueError, match="Unknown search result format"):
-            await client.chunk_repository._process_search_results(_Frame())  # ty: ignore[invalid-argument-type]
+            await client.chunk_repository._process_search_results(_Frame())

tests/test_client.py

@@ -2518,3 +2518,54 @@ async def test_visualize_chunk_falls_back_when_expansion_drops_refs(temp_db_path
             images = await client.visualize_chunk(stored[0])
 
     assert len(images) == 1
+
+
+@pytest.mark.vcr()
+async def test_import_documents_schedules_vacuum_when_enabled(temp_db_path):
+    """A batch import with auto_vacuum on queues a background vacuum."""
+    from haiku.rag.config import AppConfig
+
+    config = AppConfig()
+    config.storage.auto_vacuum = True
+
+    async with HaikuRAG(temp_db_path, config=config, create=True) as client:
+        docling_doc = await client.convert("Batch imported body.")
+        chunks = await client.chunk(docling_doc)
+        await client.import_documents(
+            [
+                DocumentImport(
+                    docling_document=docling_doc,
+                    chunks=chunks,
+                    uri="test://batch-vacuum",
+                )
+            ]
+        )
+
+        assert client._vacuum_tasks
+
+
+@pytest.mark.vcr()
+async def test_reingesting_a_source_applies_an_explicit_title(temp_db_path):
+    """Re-adding an unchanged source with a new title updates just the title."""
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        with tempfile.TemporaryDirectory() as temp_dir:
+            source = Path(temp_dir) / "retitled.txt"
+            source.write_text("stable content")
+
+            first = await client.create_document_from_source(source)
+            assert not isinstance(first, list)
+            source.write_text("changed content")
+
+            second = await client.create_document_from_source(
+                source, title="Explicit Title"
+            )
+
+        assert not isinstance(second, list)
+        assert second.id == first.id
+        assert second.title == "Explicit Title"
+
+
+async def test_update_document_rejects_unknown_id(temp_db_path):
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        with pytest.raises(ValueError, match="not found"):
+            await client.update_document("no-such-document", content="x")

tests/test_config.py

@@ -42,17 +42,17 @@ def test_find_config_file_cwd(tmp_path, monkeypatch):
 def test_find_config_file_user_config(tmp_path, monkeypatch):
     """Test finding config in user config directory."""
     monkeypatch.delenv("HAIKU_RAG_CONFIG_PATH", raising=False)
-    monkeypatch.chdir(tmp_path)
 
-    # Mock get_default_data_dir to return tmp_path
-    def mock_get_default_data_dir():
-        return tmp_path
+    # The data dir must differ from the cwd, or the cwd branch answers first
+    # and this never reaches the user-directory lookup.
+    cwd = tmp_path / "cwd"
+    cwd.mkdir()
+    data_dir = tmp_path / "data"
+    data_dir.mkdir()
+    monkeypatch.chdir(cwd)
+    monkeypatch.setattr("haiku.rag.utils.get_default_data_dir", lambda: data_dir)
 
-    monkeypatch.setattr(
-        "haiku.rag.utils.get_default_data_dir", mock_get_default_data_dir
-    )
-
-    config_file = tmp_path / "haiku.rag.yaml"
+    config_file = data_dir / "haiku.rag.yaml"
     config_file.write_text("environment: production")
 
     found = find_config_file()
@@ -518,3 +518,26 @@ def test_expand_env_var_plain_string_unchanged(tmp_path):
 
     config = load_yaml_config(config_file)
     assert config["environment"] == "production"
+
+
+def test_find_config_file_returns_none_when_nothing_exists(tmp_path, monkeypatch):
+    """No env var, no file in cwd, none in the data dir."""
+    monkeypatch.delenv("HAIKU_RAG_CONFIG_PATH", raising=False)
+    monkeypatch.chdir(tmp_path)
+
+    empty_data_dir = tmp_path / "data"
+    empty_data_dir.mkdir()
+    monkeypatch.setattr("haiku.rag.utils.get_default_data_dir", lambda: empty_data_dir)
+
+    assert find_config_file() is None
+
+
+def test_load_default_config_falls_back_to_builtin_defaults(monkeypatch):
+    """With no config file discoverable, the packaged defaults are used."""
+    from haiku.rag.config import _load_default_config
+
+    monkeypatch.setattr("haiku.rag.config.find_config_file", lambda _=None: None)
+
+    config = _load_default_config()
+
+    assert config.environment == AppConfig().environment

tests/test_context.py

@@ -1373,3 +1373,128 @@ class TestSpanInWindow:
         # A picture occupies no characters, so containment is by position.
         assert _span_in_window((10, 10, item), 0, 20) is True
         assert _span_in_window((30, 30, item), 0, 20) is False
+
+
+@pytest.mark.asyncio
+class TestExpandWithItemsWindowEdges:
+    async def test_empty_window_returns_original_results(
+        self, temp_db_path, monkeypatch
+    ):
+        """Refs resolve but the surrounding window comes back empty, so there is
+        nothing to expand from."""
+        from haiku.rag.client import HaikuRAG
+        from haiku.rag.store.models.document import Document
+
+        async with HaikuRAG(temp_db_path, create=True) as rag:
+            doc = await rag.document_repository.create(
+                Document(content="body", uri="test://window")
+            )
+            assert doc.id is not None
+            await rag.document_item_repository.create_items(
+                doc.id,
+                [
+                    DocumentItem(
+                        document_id=doc.id,
+                        position=0,
+                        self_ref="#/texts/0",
+                        label="paragraph",
+                        text="body",
+                        page_numbers=[1],
+                    )
+                ],
+            )
+
+            async def no_window(*_args, **_kwargs):
+                return []
+
+            monkeypatch.setattr(
+                rag.document_item_repository, "get_items_in_range", no_window
+            )
+
+            result = SearchResult(
+                content="original",
+                score=0.9,
+                document_id=doc.id,
+                doc_item_refs=["#/texts/0"],
+            )
+            expanded = await expand_with_items(
+                rag.document_item_repository, doc.id, [result], 5000
+            )
+
+            assert [r.content for r in expanded] == ["original"]
+
+    async def test_result_with_unmatched_refs_passes_through(self, temp_db_path):
+        """Two results share a document; the one whose refs resolve is expanded
+        and the other is returned unchanged."""
+        from haiku.rag.client import HaikuRAG
+        from haiku.rag.store.models.document import Document
+
+        async with HaikuRAG(temp_db_path, create=True) as rag:
+            doc = await rag.document_repository.create(
+                Document(content="body", uri="test://mixed")
+            )
+            assert doc.id is not None
+            await rag.document_item_repository.create_items(
+                doc.id,
+                [
+                    DocumentItem(
+                        document_id=doc.id,
+                        position=i,
+                        self_ref=f"#/texts/{i}",
+                        label="paragraph",
+                        text=f"paragraph {i}",
+                        page_numbers=[1],
+                    )
+                    for i in range(2)
+                ],
+            )
+
+            resolvable = SearchResult(
+                content="paragraph 0",
+                score=0.9,
+                document_id=doc.id,
+                doc_item_refs=["#/texts/0"],
+            )
+            unmatched = SearchResult(
+                content="untouched",
+                score=0.5,
+                document_id=doc.id,
+                doc_item_refs=["#/texts/404"],
+            )
+
+            expanded = await expand_with_items(
+                rag.document_item_repository, doc.id, [resolvable, unmatched], 5000
+            )
+
+            assert "untouched" in [r.content for r in expanded]
+
+
+def test_build_result_skips_positions_with_no_item():
+    """A sparse position map (items removed or never stored) leaves gaps in the
+    range; those positions contribute nothing."""
+    from haiku.rag.context import _build_result
+
+    original = SearchResult(content="p0", score=0.9, document_id="d1")
+    # Positions 1 and 2 in the 0..3 range carry no item.
+    pos_to_item = {
+        0: DocumentItem(
+            document_id="d1",
+            position=0,
+            self_ref="#/texts/0",
+            label="paragraph",
+            text="first",
+            page_numbers=[1],
+        ),
+        3: DocumentItem(
+            document_id="d1",
+            position=3,
+            self_ref="#/texts/3",
+            label="paragraph",
+            text="last",
+            page_numbers=[1],
+        ),
+    }
+
+    built = _build_result(0, 3, [original], pos_to_item, False, 5000)
+
+    assert built.content == "first\n\nlast"

tests/test_converters.py

@@ -1672,7 +1672,7 @@ class TestDoclingServeZipParsing:
     @pytest.mark.asyncio
     async def test_convert_text_plain_builds_document_locally(self, converter):
         """format="plain" never reaches the network."""
-        converter.client.submit_and_poll_zip = AsyncMock(  # ty: ignore[invalid-assignment]
+        converter.client.submit_and_poll_zip = AsyncMock(
             side_effect=AssertionError("must not call docling-serve")
         )
 
@@ -1680,3 +1680,18 @@ class TestDoclingServeZipParsing:
 
         assert isinstance(doc, DoclingDocument)
         assert "just text" in doc.export_to_markdown()
+
+
+@pytest.mark.asyncio
+async def test_docling_serve_convert_file_wraps_text_read_failure(tmp_path):
+    """An undecodable text file surfaces as a ValueError naming the path."""
+    config = AppConfig()
+    config.processing.converter = "docling-serve"
+    converter = get_converter(config)
+    assert isinstance(converter, DoclingServeConverter)
+
+    source = tmp_path / "broken.txt"
+    source.write_bytes(b"\xff\xfe\x00\x01 not utf-8")
+
+    with pytest.raises(ValueError, match="Failed to read text file"):
+        await converter.convert_file(source)

tests/test_database_autocreate.py

@@ -43,3 +43,26 @@ async def test_operations_work_after_database_created(tmp_path):
         doc = await client.get_document_by_id(docs[0].id)
         assert doc is not None
         assert doc.content == "Test content"
+
+
+@pytest.mark.asyncio
+async def test_default_db_path_comes_from_storage_data_dir(tmp_path):
+    """Omitting db_path places the database under the configured data dir."""
+    from haiku.rag.client import HaikuRAG
+    from haiku.rag.config import AppConfig
+
+    config = AppConfig()
+    config.storage.data_dir = tmp_path
+
+    client = HaikuRAG(config=config)
+
+    assert client._db_path == tmp_path / "haiku.rag.lancedb"
+
+
+@pytest.mark.asyncio
+async def test_vacuum_is_callable_on_the_client(temp_db_path):
+    """The public vacuum() delegates to the store."""
+    from haiku.rag.client import HaikuRAG
+
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        await client.vacuum()

tests/test_doctor.py

@@ -1206,3 +1206,26 @@ async def test_duplicate_documents_check_reads_config(temp_db_path):
         ).severity
         is Severity.WARN
     )
+
+
+@pytest.mark.asyncio
+async def test_many_unembedded_chunks_are_sampled(temp_db_path):
+    """Beyond the sample limit the detail list ends with a count of the rest."""
+    db = await _build_db(temp_db_path)
+    chunks_tbl = await db.open_table("chunks")
+    await chunks_tbl.add(
+        [
+            ChunkRecord(
+                id=f"z{i}",
+                document_id="d1",
+                content="x",
+                metadata=json.dumps({"doc_item_refs": ["#/texts/0"]}),
+                vector=[0.0] * VECTOR_DIM,
+            )
+            for i in range(8)
+        ]
+    )
+    report = await run_doctor(_config(), temp_db_path, {})
+    details = _result(report, "unembedded_chunks").details
+    assert len(details) == 6
+    assert details[-1] == "... (+3 more)"

tests/test_download_models.py

@@ -6,6 +6,7 @@ import pytest
 
 from haiku.rag.client.downloads import download_models
 from haiku.rag.config import Config
+from haiku.rag.config.models import ModelConfig
 
 
 @pytest.fixture
@@ -108,3 +109,73 @@ async def test_download_models_no_ollama_models(mock_to_thread):
     models = {e.model for e in events}
     assert "qwen3-embedding:4b" not in models
     assert "gpt-oss" not in models
+
+
+@pytest.mark.parametrize(
+    "configure,expected_model",
+    [
+        (
+            lambda c: setattr(
+                c.reranking,
+                "model",
+                ModelConfig(provider="ollama", name="rerank-model"),
+            ),
+            "rerank-model",
+        ),
+        (
+            lambda c: (
+                setattr(c.processing, "pictures", "description"),
+                setattr(
+                    c.processing.conversion_options.picture_description.model,
+                    "provider",
+                    "ollama",
+                ),
+                setattr(
+                    c.processing.conversion_options.picture_description.model,
+                    "name",
+                    "vision-model",
+                ),
+            ),
+            "vision-model",
+        ),
+        (
+            lambda c: (
+                setattr(c.processing, "auto_title", True),
+                setattr(c.processing.title_model, "provider", "ollama"),
+                setattr(c.processing.title_model, "name", "title-model"),
+            ),
+            "title-model",
+        ),
+    ],
+    ids=["reranker", "picture_description", "auto_title"],
+)
+async def test_ollama_models_from_every_config_slot_are_pulled(
+    mock_to_thread, configure, expected_model
+):
+    """Each config slot that can name an ollama model contributes to the pull set."""
+    from haiku.rag.config import AppConfig
+
+    config = AppConfig()
+    configure(config)
+
+    @asynccontextmanager
+    async def mock_stream(method, url, **kwargs):
+        mock_resp = AsyncMock()
+
+        async def aiter_lines():
+            yield '{"status": "success"}'
+
+        mock_resp.aiter_lines = aiter_lines
+        yield mock_resp
+
+    with patch(
+        "haiku.rag.client.downloads.httpx.AsyncClient",
+        return_value=_mock_httpx_client(mock_stream),
+    ):
+        pulled = {
+            progress.model
+            async for progress in download_models(config)
+            if progress.status == "pulling"
+        }
+
+    assert expected_model in pulled

tests/test_lancedb_connection.py

@@ -355,3 +355,36 @@ class TestVectorIndexCreation:
 
             indexes = await store.chunks_table.list_indices()
             assert not any("vector" in idx.columns for idx in indexes)
+
+
+class TestStoreMiscellany:
+    @pytest.mark.asyncio
+    async def test_create_makes_missing_parent_directories(self, tmp_path):
+        nested = tmp_path / "a" / "b" / "db.lancedb"
+
+        async with Store(nested, create=True) as store:
+            assert store._is_new_db is True
+
+        assert nested.exists()
+
+    @pytest.mark.asyncio
+    async def test_stored_vector_dim_is_none_for_corrupt_settings(self, temp_db_path):
+        async with Store(temp_db_path, create=True) as store:
+            await store.settings_table.update(
+                {"settings": "not json at all"}, where="id = 'settings'"
+            )
+
+            assert await store._get_stored_vector_dim() is None
+
+    @pytest.mark.asyncio
+    async def test_vacuum_skips_when_already_running(self, temp_db_path):
+        async with Store(temp_db_path, create=True) as store:
+            async with store._vacuum_lock:
+                # Returns immediately rather than blocking on the held lock.
+                await store.vacuum()
+
+    @pytest.mark.asyncio
+    async def test_history_rejects_unknown_table(self, temp_db_path):
+        async with Store(temp_db_path, create=True) as store:
+            with pytest.raises(ValueError, match="Unknown table"):
+                await store.list_table_versions("not_a_table")

tests/test_picture_in_context.py

@@ -1011,3 +1011,32 @@ async def test_rag_capability_attaches_images_for_vision_model(temp_db_path):
     assert isinstance(result, ToolReturn)
     assert result.content is not None
     assert any(isinstance(part, BinaryContent) for part in result.content)
+
+
+def test_build_picture_chunks_records_provenance_pages():
+    """A picture with provenance contributes its page numbers to the chunk."""
+    from docling_core.types.doc.base import BoundingBox
+    from docling_core.types.doc.document import ProvenanceItem
+
+    from haiku.rag.client.processing import build_picture_chunks
+    from tests.store.test_document_items import _docling_doc_with_picture
+
+    doc = _docling_doc_with_picture()
+    picture = doc.pictures[0]
+    picture.prov = [
+        ProvenanceItem(
+            page_no=3,
+            bbox=BoundingBox(l=0, t=10, r=10, b=0),
+            charspan=(0, 0),
+        ),
+        # A repeat of the same page must not be counted twice.
+        ProvenanceItem(
+            page_no=3,
+            bbox=BoundingBox(l=0, t=20, r=10, b=10),
+            charspan=(0, 0),
+        ),
+    ]
+
+    chunks = build_picture_chunks(doc, document_id="doc-1")
+
+    assert chunks[0].metadata["page_numbers"] == [3]

tests/test_processing.py

@@ -211,3 +211,56 @@ def test_merge_picture_chunks_no_pictures_returns_text_chunks():
 
     assert result is text_chunks
     assert [c.order for c in result] == [0, 1]
+
+
+@pytest.mark.asyncio
+async def test_convert_dispatches_large_pdfs_through_split_and_merge(
+    tmp_path, monkeypatch
+):
+    """With split_pages configured, PDF conversion routes through the
+    split-and-merge helper rather than the converter directly."""
+    from docling_core.types.doc.document import DoclingDocument
+
+    config = AppConfig()
+    config.processing.split_pages = 2
+
+    pdf = tmp_path / "big.pdf"
+    pdf.write_bytes(b"%PDF-1.4 stub")
+    called: dict = {}
+
+    async def fake_split(converter, path, uri, slice_size):
+        called["slice_size"] = slice_size
+        called["path"] = path
+        return DoclingDocument(name="merged")
+
+    monkeypatch.setattr(
+        "haiku.rag.converters.pdf_split.convert_pdf_with_splitting", fake_split
+    )
+
+    doc = await convert(config, pdf)
+
+    assert doc.name == "merged"
+    assert called["slice_size"] == 2
+    assert called["path"] == pdf
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "make_source,match",
+    [
+        (lambda d: (d / "missing.md").as_uri(), "File does not exist"),
+        (lambda d: _write_unsupported(d), "Unsupported file extension"),
+    ],
+    ids=["missing_file", "unsupported_extension"],
+)
+async def test_convert_rejects_bad_file_uris(tmp_path, make_source, match):
+    from haiku.rag.client.exceptions import UnsupportedSourceError
+
+    with pytest.raises(UnsupportedSourceError, match=match):
+        await convert(AppConfig(), make_source(tmp_path))
+
+
+def _write_unsupported(directory):
+    target = directory / "thing.sqlite3"
+    target.write_bytes(b"binary")
+    return target.as_uri()

tests/test_title_generation.py

@@ -371,3 +371,40 @@ class TestRebuildTitleOnly:
 
             # Only the second doc should have been processed
             assert len(processed_ids) == 1
+
+
+@pytest.mark.asyncio
+async def test_generate_title_with_llm_returns_model_output(monkeypatch):
+    """The agent's output is stripped and returned."""
+    from pydantic_ai.messages import ModelMessage, ModelResponse, TextPart
+    from pydantic_ai.models.function import AgentInfo, FunctionModel
+
+    from haiku.rag.client.titles import generate_title_with_llm
+    from haiku.rag.config import AppConfig
+
+    def respond(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
+        return ModelResponse(parts=[TextPart("  A Generated Title  ")])
+
+    monkeypatch.setattr(
+        "haiku.rag.utils.get_model", lambda *a, **kw: FunctionModel(respond)
+    )
+
+    assert await generate_title_with_llm(AppConfig(), "body") == "A Generated Title"
+
+
+@pytest.mark.asyncio
+async def test_generate_title_with_llm_returns_none_for_blank_output(monkeypatch):
+    from pydantic_ai.messages import ModelMessage, ModelResponse, TextPart
+    from pydantic_ai.models.function import AgentInfo, FunctionModel
+
+    from haiku.rag.client.titles import generate_title_with_llm
+    from haiku.rag.config import AppConfig
+
+    def respond(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
+        return ModelResponse(parts=[TextPart("   ")])
+
+    monkeypatch.setattr(
+        "haiku.rag.utils.get_model", lambda *a, **kw: FunctionModel(respond)
+    )
+
+    assert await generate_title_with_llm(AppConfig(), "body") is None

tests/tools/test_document.py

@@ -187,6 +187,33 @@ class TestSummarizeDocumentTool:
 
         assert "Document not found" in result
 
+    @pytest.mark.vcr()
+    @pytest.mark.asyncio
+    async def test_summarize_document_returns_model_summary(
+        self, doc_client, doc_config, monkeypatch
+    ):
+        """A resolvable document is summarised and labelled with its title."""
+        from pydantic_ai.messages import ModelMessage, ModelResponse, TextPart
+        from pydantic_ai.models.function import AgentInfo, FunctionModel
+
+        def respond(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
+            return ModelResponse(parts=[TextPart("A concise summary.")])
+
+        monkeypatch.setattr(
+            "haiku.rag.tools.document.get_model",
+            lambda *a, **kw: FunctionModel(respond),
+        )
+
+        docs = await doc_client.list_documents()
+        assert docs and docs[0].uri
+
+        toolset = create_document_toolset(doc_config)
+        summarize_tool = toolset.tools["summarize_document"]
+        result = await summarize_tool.function(make_ctx(doc_client), docs[0].uri)
+
+        assert "A concise summary." in result
+        assert "Summary of" in result
+
 
 @pytest.fixture
 async def doc_client(temp_db_path):