fix convert() misreading text content that starts with a URL

CHANGELOG.md

@@ -1,6 +1,10 @@
 # Changelog
 ## [Unreleased]
 
+### Fixed
+
+- **`create_document`, `update_document`, and rebuild (`RECHUNK` / full fallback) no longer misread URL-prefixed text as a URL to fetch.** These paths passed known-text content through `HaikuRAG.convert()`, which dispatches on `urlparse(source).scheme`; text whose first line was `https://...` (common for clipped web pages and notes) got handed to `httpx.get` and crashed with `httpx.InvalidURL` on embedded whitespace. Fixed by calling `converter.convert_text(...)` directly at those sites; `convert()` itself is unchanged for `create_document_from_source`.
+
 ### Changed
 
 - **Skills share a single `HaikuRAG` client per invocation** via the new `haiku.skills>=0.15.0` `lifespan` hook. The skill's sub-agent opens one read-only client on entry, all tool calls reuse it, and it closes on exit — replacing the old pattern of open/close around every `search` / `list_documents` / `get_document` call.

haiku_rag_slim/haiku/rag/client.py

@@ -495,7 +495,8 @@ class HaikuRAG:
         from haiku.rag.embeddings import embed_chunks
 
         # Convert → Chunk → Embed using primitives
-        docling_document = await self.convert(content, format=format)
+        converter = get_converter(self._config)
+        docling_document = await converter.convert_text(content, format=format)
         chunks = await self.chunk(docling_document)
         embedded_chunks = await embed_chunks(chunks, self._config)
 
@@ -1000,7 +1001,10 @@ class HaikuRAG:
         # Content provided without chunks - convert, chunk, and embed using primitives
         assert content is not None
         existing_doc.content = content
-        converted_docling = await self.convert(existing_doc.content)
+        converter = get_converter(self._config)
+        converted_docling = await converter.convert_text(
+            existing_doc.content, format="md"
+        )
         existing_doc.set_docling(converted_docling)
 
         new_chunks = await self.chunk(converted_docling)
@@ -1558,11 +1562,13 @@ class HaikuRAG:
         pending_docs: list[Document] = []
         pending_doc_ids: list[str] = []
 
+        converter = get_converter(self._config)
+
         for doc in documents:
             assert doc.id is not None
 
-            # Convert content to DoclingDocument
-            docling_document = await self.convert(doc.content)
+            # Convert stored markdown to DoclingDocument
+            docling_document = await converter.convert_text(doc.content, format="md")
 
             # Chunk and embed
             chunks = await self.chunk(docling_document)
@@ -1605,6 +1611,7 @@ class HaikuRAG:
         pending_chunks: list[Chunk] = []
         pending_docs: list[Document] = []
         pending_doc_ids: list[str] = []
+        converter = get_converter(self._config)
 
         for doc in documents:
             assert doc.id is not None
@@ -1643,7 +1650,7 @@ class HaikuRAG:
                     "Source missing for %s, re-embedding from content", doc.uri
                 )
 
-            docling_document = await self.convert(doc.content)
+            docling_document = await converter.convert_text(doc.content, format="md")
             chunks = await self.chunk(docling_document)
             embedded_chunks = await embed_chunks(chunks, self._config)
 

tests/test_client.py

@@ -1544,3 +1544,78 @@ async def test_sql_injection_is_blocked_with_escaping(temp_db_path):
             filter=f"title = '{injection_payload}'"
         )
         assert len(docs_unescaped) == 2  # SQL injection succeeds without escaping
+
+
+# =============================================================================
+# URL-prefixed content regression tests
+# =============================================================================
+
+
+def _patch_embed_chunks(monkeypatch):
+    async def fake_embed_chunks(chunks, config):
+        for chunk in chunks:
+            chunk.embedding = [0.0] * 2560
+        return chunks
+
+    monkeypatch.setattr("haiku.rag.embeddings.embed_chunks", fake_embed_chunks)
+
+
+async def test_create_document_with_url_prefixed_content(temp_db_path, monkeypatch):
+    """Text whose first line is a URL must be stored as text, not fetched."""
+    _patch_embed_chunks(monkeypatch)
+
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        content = "https://example.com/foo\n\n# Heading\n\nBody text here."
+        doc = await client.create_document(content=content, uri="test://url-prefixed")
+
+        assert doc.id is not None
+        assert "example.com" in doc.content
+        assert "Heading" in doc.content
+
+
+async def test_update_document_with_url_prefixed_content(temp_db_path, monkeypatch):
+    """update_document(content=...) with URL-prefixed text must not fetch it."""
+    _patch_embed_chunks(monkeypatch)
+
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        doc = await client.create_document(
+            content="initial body", uri="test://update-url"
+        )
+        assert doc.id is not None
+
+        url_prefixed = "https://example.com/bar\n\n# New heading\n\nReplacement body."
+        updated = await client.update_document(doc.id, content=url_prefixed)
+
+        assert "example.com" in updated.content
+        assert "New heading" in updated.content
+
+
+async def test_rebuild_rechunk_with_url_prefixed_stored_content(
+    temp_db_path, monkeypatch
+):
+    """RECHUNK rebuild must handle stored markdown whose first line is a URL."""
+    from haiku.rag.client import RebuildMode
+
+    _patch_embed_chunks(monkeypatch)
+
+    async with HaikuRAG(temp_db_path, create=True) as client:
+        doc = await client.create_document(
+            content="plain seed content", uri="file:///nonexistent/path.txt"
+        )
+        assert doc.id is not None
+
+        # Overwrite stored content to simulate markdown that starts with a URL,
+        # bypassing the (also-affected) create_document path so this test
+        # specifically exercises the rebuild path.
+        doc.content = "https://example.com/baz\n\n# Stored\n\nStored body text."
+        await client.document_repository.update(doc)
+
+        processed_ids = [
+            doc_id async for doc_id in client.rebuild_database(mode=RebuildMode.RECHUNK)
+        ]
+        assert doc.id in processed_ids
+
+        doc_after = await client.document_repository.get_by_id(doc.id)
+        assert doc_after is not None
+        assert "example.com" in doc_after.content
+        assert "Stored" in doc_after.content