Tidy PDF attachment extension fix

CHANGELOG.md

@@ -1,6 +1,10 @@
 # Changelog
 ## [Unreleased]
 
+### Fixed
+
+- Embedded PDF attachment extension is derived from the attachment filename, not the parent's synthetic `...#attachment=<name>` URI; non-PDF attachments (e.g. `.joboptions`) are no longer misrouted to docling's PDF backend, and unsupported extensions are skipped.
+
 ## [0.57.0] - 2026-06-11
 
 ### Added

haiku_rag_slim/haiku/rag/client/documents.py

@@ -338,15 +338,10 @@ async def _ingest_fetch_result(
     existing document if one is supplied. ``depth`` tracks position in an
     attachment chain so the reconciliation step can bound recursion.
 
-    ``filename`` makes that name's suffix the authoritative source of the
+    ``filename``, when given, makes its suffix authoritative for the file
-    file extension (hence the docling format). Callers pass it when the
+    extension (and thus the docling format), overriding the URI/content-type
-    result's ``uri`` cannot yield the right extension -- notably embedded
+    fallback. Callers pass it when ``result.uri`` cannot yield the right
-    PDF attachments, whose synthetic ``...#attachment=<name>`` URI carries
+    extension, e.g. embedded attachments whose name lives in a URI fragment."""
-    the attachment name in the fragment, so the URL-suffix fallback would
-    otherwise inherit the *parent* PDF's ``.pdf`` and feed ASCII payloads
-    (e.g. ``Press Quality.joboptions``) to docling's PDF backend. An empty
-    or unsupported suffix then fails the guard below and is rejected,
-    instead of being silently misrouted."""
     from haiku.rag.embeddings import embed_chunks
 
     converter = get_converter(client._config)

tests/test_pdf_attachments.py

@@ -332,42 +332,21 @@ async def test_unsupported_attachment_continues_loop(temp_db_path, monkeypatch):
 
 
 async def test_joboptions_attachment_skipped_not_routed_as_pdf(temp_db_path, caplog):
-    """Regression: an Adobe ``.joboptions`` preset is ASCII text, embedded by
+    """An attachment's own name drives its extension, not the synthetic
-    name only. Its synthetic ``...#attachment=Press%20Quality.joboptions`` URI
+    ``...#attachment=<name>`` URI (whose fragment the URL-suffix fallback drops,
-    carries the name in a *fragment*, so the URL-suffix fallback used to inherit
+    inheriting the parent's ``.pdf``). ``.joboptions`` is unsupported, so the
-    the PARENT PDF's ``.pdf`` and hand the ASCII bytes to docling's PDF backend.
+    child is skipped before any converter/embedder call (real ingest, no fake)."""
-
-    The attachment's own name must now drive the extension: ``.joboptions`` is
-    unsupported, so the child is skipped -- never ingested, never sent to
-    docling. This runs the REAL ``_ingest_fetch_result`` (no monkeypatch); its
-    unsupported-extension guard short-circuits before any converter/embedder
-    call, so the sole attachment needs no LLM."""
     import logging
 
-    from haiku.rag.client.processing import get_extension_from_content_type_or_url
-
     pdf_bytes = build_pdf([("Press Quality.joboptions", b"/CompressObjects /Tags\n")])
     async with HaikuRAG(temp_db_path, create=True) as client:
         parent_uri = "file:///fixtures/brochure.pdf"
         parent = await _make_parent(client, parent_uri, pdf_bytes)
 
-        # The latent trap: octet-stream + the fragment URI still resolves to the
-        # parent's ``.pdf`` -- which is exactly what the fix must no longer use.
-        child_uri = f"{parent_uri}#attachment=Press%20Quality.joboptions"
-        assert (
-            get_extension_from_content_type_or_url(
-                child_uri, "application/octet-stream"
-            )
-            == ".pdf"
-        )
-
         with caplog.at_level(logging.WARNING, logger="haiku.rag.client.documents"):
             await _reconcile_pdf_attachments(client, parent, pdf_bytes, depth=0)
 
-        # No child created: the ASCII preset was skipped, not parsed as a PDF.
         assert await client.list_documents(filter=parent_uri_filter(parent_uri)) == []
-        assert "unsupported extension" in caplog.text
-        assert ".joboptions" in caplog.text
 
 
 async def test_cascade_delete_removes_reconciled_children(temp_db_path, monkeypatch):