fix(images): write note-relative image links in sources pages

openkb/agent/query.py

@@ -37,8 +37,11 @@
    - PageIndex documents (doc_type: pageindex): use get_page_content(doc_name, pages)
      with tight page ranges. The summary shows document tree structure with page
      ranges to help you target. Never fetch the whole document.
-6. Source content may reference images (e.g. ![image](sources/images/doc/file.png)).
-   Use the get_image tool to view them when needed.
+6. Source content may reference images. Short-doc .md pages link them
+   note-relative (e.g. ![image](images/doc/file.png), resolved from
+   wiki/sources/); long-doc JSON page metadata lists them wiki-root-relative
+   (e.g. sources/images/doc/file.png). Pass either form as seen to the
+   get_image tool — it accepts both.
 7. Synthesize a clear, concise, well-cited answer grounded in wiki content.
 
 Answer based only on wiki content. Be concise.
@@ -81,7 +84,10 @@ def get_image(image_path: str) -> ToolOutputImage | ToolOutputText:
         you'd need to see to answer accurately.
 
         Args:
-            image_path: Image path relative to wiki root (e.g. 'sources/images/doc/p1_img1.png').
+            image_path: Image path as it appears in the content — either
+                wiki-root-relative ('sources/images/doc/p1_img1.png') or
+                note-relative as used in sources/ .md pages
+                ('images/doc/p1_img1.png').
         """
         result = read_wiki_image(image_path, wiki_root)
         if result["type"] == "image":

openkb/agent/tools.py

@@ -147,7 +147,10 @@ def read_wiki_image(path: str, wiki_root: str) -> dict:
     """Read an image file from the wiki and return as base64 data URL.
 
     Args:
-        path: Image path relative to *wiki_root* (e.g. ``"sources/images/doc/p1_img1.png"``).
+        path: Image path relative to *wiki_root*
+            (e.g. ``"sources/images/doc/p1_img1.png"``), or note-relative
+            as embedded in sources/ .md pages
+            (``"images/doc/p1_img1.png"`` — retried under ``sources/``).
         wiki_root: Absolute path to the wiki root directory.
 
     Returns:
@@ -161,7 +164,13 @@ def read_wiki_image(path: str, wiki_root: str) -> dict:
     if not full_path.is_relative_to(root):
         return {"type": "text", "text": "Access denied: path escapes wiki root."}
     if not full_path.exists():
-        return {"type": "text", "text": f"Image not found: {path}"}
+        # Source .md pages embed images note-relative ("images/<doc>/<file>",
+        # resolved from wiki/sources/). Callers pass those verbatim — retry
+        # under sources/ before failing.
+        alt_path = (root / "sources" / path).resolve()
+        if not (alt_path.is_relative_to(root) and alt_path.exists()):
+            return {"type": "text", "text": f"Image not found: {path}"}
+        full_path = alt_path
 
     mime = _MIME_TYPES.get(full_path.suffix.lower(), "image/png")
     b64 = base64.b64encode(full_path.read_bytes()).decode()

openkb/images.py

@@ -23,6 +23,19 @@
 _MIN_IMAGE_DIM = 32
 
 
+def md_image_ref(alt: str, doc_name: str, filename: str) -> str:
+    """Markdown image reference as written into ``wiki/sources/<doc>.md``.
+
+    Note-relative (``images/{doc}/{file}``): source pages live directly in
+    ``wiki/sources/``, so this resolves in every renderer that resolves links
+    relative to the containing file (Obsidian, GitHub, VS Code). Internal
+    metadata (the per-page JSON of long docs) keeps wiki-root-relative
+    ``sources/images/...`` paths instead — those are consumed by tools that
+    resolve against the wiki root, not rendered from a note.
+    """
+    return f"![{alt}](images/{doc_name}/{filename})"
+
+
 def extract_pdf_images(pdf_path: Path, doc_name: str, images_dir: Path) -> dict[int, list[str]]:
     """Extract images from a PDF using pymupdf's dict-mode block iteration.
 
@@ -31,7 +44,9 @@ def extract_pdf_images(pdf_path: Path, doc_name: str, images_dir: Path) -> dict[
     as PNG. This captures both embedded bitmaps *and* vector-rendered figures
     that ``get_images()`` would miss.
 
-    Returns a mapping of page_number (1-based) → list of relative image paths.
+    Returns a mapping of page_number (1-based) → list of image paths. Paths
+    are wiki-root-relative (``sources/images/...``) — internal metadata for
+    tools that resolve against the wiki root, not note-rendered markdown.
     """
     images_dir.mkdir(parents=True, exist_ok=True)
     page_images: dict[int, list[str]] = {}
@@ -77,7 +92,10 @@ def convert_pdf_to_pages(pdf_path: Path, doc_name: str, images_dir: Path) -> lis
     """Convert a PDF to per-page dicts with text content and images.
 
     Each dict has ``{"page": int, "content": str, "images": [{"path": str}]}``.
-    Images are saved to *images_dir* and referenced with wiki-root-relative paths.
+    Images are saved to *images_dir* and referenced with wiki-root-relative
+    ``sources/images/...`` paths — these pages land in ``sources/<doc>.json``
+    (never rendered from a note), and both ``get_wiki_page_content`` and
+    ``read_wiki_image`` resolve them against the wiki root.
     """
     images_dir.mkdir(parents=True, exist_ok=True)
     pages: list[dict] = []
@@ -134,8 +152,9 @@ def convert_pdf_with_images(pdf_path: Path, doc_name: str, images_dir: Path) ->
     """Convert a PDF to markdown with inline images using pymupdf dict-mode.
 
     Iterates blocks in reading order per page. Text blocks become text,
-    image blocks are saved to disk and replaced with ``![image](path)``
-    inline — preserving the original position in the document.
+    image blocks are saved to disk and replaced with a note-relative
+    ``![image](images/{doc_name}/...)`` link inline — preserving the
+    original position in the document.
 
     Returns the full markdown string.
     """
@@ -173,7 +192,7 @@ def convert_pdf_with_images(pdf_path: Path, doc_name: str, images_dir: Path) ->
                         filename = f"p{page_num}_img{img_counter}.png"
                         (images_dir / filename).write_bytes(pix.tobytes("png"))
                         pix = None
-                        parts.append(f"\n![image](sources/images/{doc_name}/{filename})\n")
+                        parts.append(f"\n{md_image_ref('image', doc_name, filename)}\n")
                     except Exception:
                         logger.warning("Failed to save image block on page %d", page_num)
     return "\n".join(parts)
@@ -184,7 +203,7 @@ def extract_base64_images(markdown: str, doc_name: str, images_dir: Path) -> str
 
     For each ``![alt](data:image/ext;base64,DATA)`` match:
     - Decode base64 bytes → save to ``images_dir/img_NNN.ext``
-    - Replace the link with ``![alt](sources/images/{doc_name}/img_NNN.ext)``
+    - Replace the link with ``![alt](images/{doc_name}/img_NNN.ext)``
     - On decode failure: log a warning and leave the original text unchanged.
     """
     counter = 0
@@ -208,7 +227,7 @@ def extract_base64_images(markdown: str, doc_name: str, images_dir: Path) -> str
         images_dir.mkdir(parents=True, exist_ok=True)
         dest.write_bytes(image_bytes)
 
-        new_ref = f"![{alt}](sources/images/{doc_name}/{filename})"
+        new_ref = md_image_ref(alt, doc_name, filename)
         result = result.replace(match.group(0), new_ref, 1)
 
     return result
@@ -220,7 +239,7 @@ def copy_relative_images(markdown: str, source_dir: Path, doc_name: str, images_
     For each ``![alt](relative/path)`` match (skipping http/https and data URIs):
     - Resolve path relative to ``source_dir``
     - Copy to ``images_dir/{filename}``
-    - Replace link with ``![alt](sources/images/{doc_name}/{filename})``
+    - Replace link with ``![alt](images/{doc_name}/{filename})``
     - Missing source file: log a warning and leave the original text unchanged.
     """
     result = markdown
@@ -254,7 +273,7 @@ def copy_relative_images(markdown: str, source_dir: Path, doc_name: str, images_
             images_dir.mkdir(parents=True, exist_ok=True)
             shutil.copy2(src, images_dir / filename)
 
-        new_ref = f"![{alt}](sources/images/{doc_name}/{filename})"
+        new_ref = md_image_ref(alt, doc_name, filename)
         result = result.replace(match.group(0), new_ref, 1)
 
     return result

openkb/skill/creator.py

@@ -105,8 +105,10 @@ def get_image(image_path: str) -> ToolOutputImage | ToolOutputText:
         need to see in order to distil it correctly into the skill.
 
         Args:
-            image_path: Path relative to wiki/
-                (e.g. ``"sources/images/doc/p1_img1.png"``).
+            image_path: Image path as it appears in the content — either
+                wiki-root-relative (``"sources/images/doc/p1_img1.png"``)
+                or note-relative as used in sources/ .md pages
+                (``"images/doc/p1_img1.png"``).
         """
         result = _read_image_impl(image_path, wiki_root)
         if result["type"] == "image":

skills/openkb/references/wiki-schema.md

@@ -99,8 +99,11 @@ thing, read the matching entity page first.
 
 ## `wiki/sources/<doc>.md` (short docs)
 
-The markitdown-converted full text. Image refs appear as
-`![](sources/images/<doc>/p1_img1.png)`.
+The markitdown-converted full text. Image refs are note-relative —
+`![](images/<doc>/p1_img1.png)` — and resolve from `wiki/sources/`
+(the files live in `wiki/sources/images/<doc>/`). The per-page JSON of
+long docs instead lists images wiki-root-relative
+(`sources/images/<doc>/...`).
 
 ## `wiki/sources/<doc>.json` (long PDFs)
 

tests/test_agent_tools.py

@@ -7,9 +7,59 @@
     list_wiki_files,
     parse_pages,
     read_wiki_file,
+    read_wiki_image,
     write_wiki_file,
 )
 
+FAKE_PNG = b"\x89PNG\r\n\x1a\n" + b"\x00" * 8
+
+
+# ---------------------------------------------------------------------------
+# read_wiki_image
+# ---------------------------------------------------------------------------
+
+
+class TestReadWikiImage:
+    def _make_image(self, tmp_path):
+        images_dir = tmp_path / "sources" / "images" / "doc"
+        images_dir.mkdir(parents=True)
+        (images_dir / "p1_img1.png").write_bytes(FAKE_PNG)
+
+    def test_reads_wiki_root_relative_path(self, tmp_path):
+        self._make_image(tmp_path)
+
+        result = read_wiki_image("sources/images/doc/p1_img1.png", str(tmp_path))
+
+        assert result["type"] == "image"
+        assert result["image_url"].startswith("data:image/png;base64,")
+
+    def test_reads_note_relative_path_from_sources(self, tmp_path):
+        # Source .md pages embed images as "images/<doc>/<file>" (relative to
+        # wiki/sources/); the tool must resolve those verbatim too.
+        self._make_image(tmp_path)
+
+        result = read_wiki_image("images/doc/p1_img1.png", str(tmp_path))
+
+        assert result["type"] == "image"
+        assert result["image_url"].startswith("data:image/png;base64,")
+
+    def test_missing_image_reports_not_found(self, tmp_path):
+        self._make_image(tmp_path)
+
+        result = read_wiki_image("images/doc/nope.png", str(tmp_path))
+
+        assert result["type"] == "text"
+        assert "not found" in result["text"].lower()
+
+    def test_path_escape_denied(self, tmp_path):
+        self._make_image(tmp_path)
+
+        result = read_wiki_image("../outside.png", str(tmp_path))
+
+        assert result["type"] == "text"
+        assert "Access denied" in result["text"]
+
+
 # ---------------------------------------------------------------------------
 # list_wiki_files
 # ---------------------------------------------------------------------------

tests/test_images.py

@@ -41,7 +41,7 @@ def test_single_base64_image_extracted(self, tmp_path):
 
         # Result should reference a saved file, not the raw base64
         assert "data:image/png;base64," not in result
-        assert "![alt text](sources/images/doc/img_001.png)" == result
+        assert "![alt text](images/doc/img_001.png)" == result
 
         # File should exist on disk
         saved = images_dir / "img_001.png"
@@ -56,8 +56,8 @@ def test_multiple_base64_images_numbered_sequentially(self, tmp_path):
         md = f"![fig1](data:image/png;base64,{b64_png})\n![fig2](data:image/jpeg;base64,{b64_jpg})"
         result = extract_base64_images(md, "doc", images_dir)
 
-        assert "![fig1](sources/images/doc/img_001.png)" in result
-        assert "![fig2](sources/images/doc/img_002.jpeg)" in result
+        assert "![fig1](images/doc/img_001.png)" in result
+        assert "![fig2](images/doc/img_002.jpeg)" in result
         assert (images_dir / "img_001.png").exists()
         assert (images_dir / "img_002.jpeg").exists()
 
@@ -85,7 +85,7 @@ def test_mixed_valid_invalid_base64(self, tmp_path, caplog):
 
         with caplog.at_level(logging.WARNING, logger="openkb.images"):
             result = extract_base64_images(md, "doc", images_dir)
-        assert "![good](sources/images/doc/img_001.png)" in result
+        assert "![good](images/doc/img_001.png)" in result
         assert f"data:image/png;base64,{bad}" in result
 
 
@@ -107,7 +107,7 @@ def test_existing_relative_image_copied_and_rewritten(self, tmp_path):
         md = "![diagram](diagram.png)"
         result = copy_relative_images(md, source_dir, "doc", images_dir)
 
-        assert "![diagram](sources/images/doc/diagram.png)" == result
+        assert "![diagram](images/doc/diagram.png)" == result
         assert (images_dir / "diagram.png").read_bytes() == FAKE_PNG
 
     def test_missing_relative_image_leaves_original(self, tmp_path, caplog):
@@ -157,8 +157,8 @@ def test_multiple_relative_images_all_copied(self, tmp_path):
         md = "![a](a.png)\n![b](b.jpg)"
         result = copy_relative_images(md, source_dir, "doc", images_dir)
 
-        assert "![a](sources/images/doc/a.png)" in result
-        assert "![b](sources/images/doc/b.jpg)" in result
+        assert "![a](images/doc/a.png)" in result
+        assert "![b](images/doc/b.jpg)" in result
         assert (images_dir / "a.png").exists()
         assert (images_dir / "b.jpg").exists()
 
@@ -195,4 +195,55 @@ def test_same_image_referenced_twice_is_copied_once(self, tmp_path):
         result = copy_relative_images(md, source_dir, "doc", images_dir)
 
         assert [p.name for p in images_dir.iterdir()] == ["logo.png"]
-        assert result.count("sources/images/doc/logo.png") == 2
+        assert result.count("images/doc/logo.png") == 2
+
+
+# ---------------------------------------------------------------------------
+# Note-relative resolution (Obsidian / GitHub compatibility)
+# ---------------------------------------------------------------------------
+
+
+class TestNoteRelativeResolution:
+    """Generated ``![...](...)`` links must resolve relative to the note.
+
+    Source pages live at ``wiki/sources/<doc>.md`` and their images at
+    ``wiki/sources/images/<doc>/`` — renderers that resolve links relative
+    to the containing file (Obsidian, GitHub, VS Code) must find the image.
+    A wiki-root-relative link (``sources/images/...``) would resolve to the
+    non-existent ``wiki/sources/sources/images/...`` from those notes.
+    """
+
+    @staticmethod
+    def _link_paths(markdown: str) -> list[str]:
+        import re
+
+        return re.findall(r"!\[[^\]]*\]\(([^)]+)\)", markdown)
+
+    def test_base64_image_ref_resolves_from_sources_note(self, tmp_path):
+        wiki = tmp_path / "wiki"
+        note = wiki / "sources" / "doc.md"
+        images_dir = wiki / "sources" / "images" / "doc"
+        images_dir.mkdir(parents=True)
+
+        md = f"![alt](data:image/png;base64,{_make_b64(FAKE_PNG)})"
+        result = extract_base64_images(md, "doc", images_dir)
+        note.write_text(result, encoding="utf-8")
+
+        for rel in self._link_paths(result):
+            assert (note.parent / rel).exists(), rel
+
+    def test_copied_image_ref_resolves_from_sources_note(self, tmp_path):
+        source_dir = tmp_path / "clip"
+        source_dir.mkdir()
+        (source_dir / "figure.png").write_bytes(FAKE_PNG)
+
+        wiki = tmp_path / "wiki"
+        note = wiki / "sources" / "doc.md"
+        images_dir = wiki / "sources" / "images" / "doc"
+        images_dir.mkdir(parents=True)
+
+        result = copy_relative_images("![f](figure.png)", source_dir, "doc", images_dir)
+        note.write_text(result, encoding="utf-8")
+
+        for rel in self._link_paths(result):
+            assert (note.parent / rel).exists(), rel