databricks-tellr-app 0.3.13.dev3__tar.gz → 0.3.13.dev11__tar.gz

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: databricks-tellr-app
-Version: 0.3.13.dev3
+Version: 0.3.13.dev11
 Summary: Tellr application package for Databricks Apps
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/databricks_tellr_app.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: databricks-tellr-app
-Version: 0.3.13.dev3
+Version: 0.3.13.dev11
 Summary: Tellr application package for Databricks Apps
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/databricks_tellr_app.egg-info/SOURCES.txt

@@ -115,13 +115,6 @@ setup.py
 ./src/services/profile_service.py
 ./src/services/streaming_callback.py
 ./src/services/validator.py
-./src/services/converter_jail/__init__.py
-./src/services/converter_jail/ast_guard.py
-./src/services/converter_jail/codeprep.py
-./src/services/converter_jail/gslides_runner.py
-./src/services/converter_jail/jail.py
-./src/services/converter_jail/pptx_runner.py
-./src/services/converter_jail/protocol.py
 ./src/services/evaluation/__init__.py
 ./src/services/evaluation/llm_judge.py
 ./src/services/identity_providers/__init__.py
@@ -265,13 +258,6 @@ src/services/pptx_prompts_defaults.py
 src/services/profile_service.py
 src/services/streaming_callback.py
 src/services/validator.py
-src/services/converter_jail/__init__.py
-src/services/converter_jail/ast_guard.py
-src/services/converter_jail/codeprep.py
-src/services/converter_jail/gslides_runner.py
-src/services/converter_jail/jail.py
-src/services/converter_jail/pptx_runner.py
-src/services/converter_jail/protocol.py
 src/services/evaluation/__init__.py
 src/services/evaluation/llm_judge.py
 src/services/identity_providers/__init__.py

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "databricks-tellr-app"
-version = "0.3.13.dev3"
+version = "0.3.13.dev11"
 description = "Tellr application package for Databricks Apps"
 readme = "README.md"
 requires-python = ">=3.10"

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/src/api/services/chat_service.py

@@ -80,6 +80,11 @@ class ChatService:
         # This avoids re-parsing HTML on every request
         self._deck_cache: Dict[str, SlideDeck] = {}
 
+        # DB version each cached deck corresponds to. The cache is per-process
+        # while prod runs multiple uvicorn workers sharing one database, so a
+        # cache hit is only valid if the DB version hasn't moved on.
+        self._deck_cache_versions: Dict[str, int] = {}
+
         logger.info("ChatService initialized successfully")
 
     def _substitute_images_for_response(self, deck_dict, raw_html=None):
@@ -600,7 +605,7 @@ class ChatService:
                     slide_deck_dict = current_deck.to_dict()
 
                 try:
-                    session_manager.save_slide_deck(
+                    save_result = session_manager.save_slide_deck(
                         session_id=session_id,
                         title=current_deck.title,
                         html_content=current_deck.knit(),
@@ -610,6 +615,7 @@ class ChatService:
                         modified_by=_user,
                         expected_version=_deck_version_before_llm,
                     )
+                    self._record_deck_version(session_id, save_result)
                 except VersionConflictError:
                     logger.warning(
                         "Chat save rejected: deck was edited during LLM call, reloading",
@@ -1385,7 +1391,7 @@ class ChatService:
                 slide_deck_dict = current_deck.to_dict()
 
             try:
-                session_manager.save_slide_deck(
+                save_result = session_manager.save_slide_deck(
                     session_id=session_id,
                     title=current_deck.title,
                     html_content=current_deck.knit(),
@@ -1395,6 +1401,7 @@ class ChatService:
                     modified_by=_user,
                     expected_version=_deck_version_before_llm,
                 )
+                self._record_deck_version(session_id, save_result)
             except VersionConflictError:
                 logger.warning(
                     "Chat save rejected: deck was edited during LLM call, reloading",
@@ -1921,18 +1928,63 @@ class ChatService:
         context = "\n\n[Attached images]\n" + "\n".join(image_descriptions)
         return message + context
 
+    def _deck_versions(self) -> Dict[str, int]:
+        """Return the cached-deck version map, creating it if needed.
+
+        Lazy creation keeps instances built without __init__ (test harnesses
+        use ChatService.__new__) working. Callers must hold _cache_lock.
+        """
+        try:
+            return self._deck_cache_versions
+        except AttributeError:
+            self._deck_cache_versions = {}
+            return self._deck_cache_versions
+
+    def _record_deck_version(self, session_id: str, save_result: Optional[Dict[str, Any]]) -> None:
+        """Remember which DB version the cached deck now corresponds to.
+
+        Call after save_slide_deck so the next _get_or_load_deck cache hit
+        validates cleanly instead of reloading from the database.
+        """
+        version = (save_result or {}).get("version")
+        with self._cache_lock:
+            if version is not None:
+                self._deck_versions()[session_id] = version
+            else:
+                self._deck_versions().pop(session_id, None)
+
     def _get_or_load_deck(self, session_id: str) -> Optional[SlideDeck]:
         """Get deck from cache or load from database.
 
         Thread-safe access to deck cache using _cache_lock.
-        
+
+        The cache is per-process while prod runs multiple uvicorn workers
+        sharing one database, so a cached deck is only served if its recorded
+        version matches the current DB version; otherwise it is reloaded.
+
         Uses deck_dict (slides array with individual scripts) when available,
         falling back to from_html_string for legacy data.
         """
         # Check cache first (with lock)
         with self._cache_lock:
-            if session_id in self._deck_cache:
-                return self._deck_cache[session_id]
+            cached_deck = self._deck_cache.get(session_id)
+            cached_version = self._deck_versions().get(session_id)
+
+        if cached_deck is not None:
+            db_version = self._get_deck_version(session_id)
+            if db_version is None or cached_version == db_version:
+                return cached_deck
+            logger.info(
+                "Deck cache stale, reloading from database",
+                extra={
+                    "session_id": session_id,
+                    "cached_version": cached_version,
+                    "db_version": db_version,
+                },
+            )
+            with self._cache_lock:
+                self._deck_cache.pop(session_id, None)
+                self._deck_versions().pop(session_id, None)
 
         # Try to load from database (outside lock to avoid blocking)
         session_manager = get_session_manager()
@@ -1966,6 +2018,10 @@ class ChatService:
             # Store in cache (with lock)
             with self._cache_lock:
                 self._deck_cache[session_id] = deck
+                if deck_data.get("version") is not None:
+                    self._deck_versions()[session_id] = deck_data["version"]
+                else:
+                    self._deck_versions().pop(session_id, None)
             return deck
         except Exception as e:
             logger.warning(f"Failed to load deck from database: {e}")
@@ -1975,23 +2031,18 @@ class ChatService:
     def _get_deck_version(self, session_id: str) -> Optional[int]:
         """Read the current deck version from the database.
 
-        Called once before the LLM runs to capture the version for
-        optimistic locking. Uses the existing get_slide_deck which
-        returns the full deck dict including the version column.
-
-        On Lakebase/PostgreSQL this is a single-row indexed lookup (~1-2ms).
+        Used on every deck-cache hit (multi-worker cache validation) and
+        before LLM calls (optimistic locking), so it must stay cheap: a
+        single-column indexed lookup, never a full deck fetch.
 
         Returns:
             Current deck version number, or None if no deck exists.
         """
         session_manager = get_session_manager()
         try:
-            deck_data = session_manager.get_slide_deck(session_id)
-            if deck_data:
-                return deck_data.get("version")
+            return session_manager.get_slide_deck_version(session_id)
         except Exception:
-            pass
-        return None
+            return None
 
     @staticmethod
     def _reindex_slide_ids(deck: "SlideDeck") -> None:
@@ -2008,6 +2059,7 @@ class ChatService:
         """Remove the cached deck for a session so the next read hits the DB."""
         with self._cache_lock:
             self._deck_cache.pop(session_id, None)
+            self._deck_versions().pop(session_id, None)
 
     def _replace_slide_htmls_from_cache(self, session_id: str, slide_context: Dict[str, Any]) -> Dict[str, Any]:
         """Replace frontend-supplied slide_htmls with backend cache versions.
@@ -2540,7 +2592,7 @@ class ChatService:
         # Persist to database
         deck_dict = current_deck.to_dict()
         session_manager = get_session_manager()
-        session_manager.save_slide_deck(
+        save_result = session_manager.save_slide_deck(
             session_id=session_id,
             title=current_deck.title,
             html_content=current_deck.knit(),
@@ -2549,6 +2601,7 @@ class ChatService:
             deck_dict=deck_dict,
             expected_version=expected_version,
         )
+        self._record_deck_version(session_id, save_result)
 
         # Create save point
         try:
@@ -2616,7 +2669,7 @@ class ChatService:
         # Persist to database
         deck_dict = current_deck.to_dict()
         session_manager = get_session_manager()
-        session_manager.save_slide_deck(
+        save_result = session_manager.save_slide_deck(
             session_id=session_id,
             title=current_deck.title,
             html_content=current_deck.knit(),
@@ -2625,6 +2678,7 @@ class ChatService:
             deck_dict=deck_dict,
             expected_version=expected_version,
         )
+        self._record_deck_version(session_id, save_result)
 
         # Create save point immediately after persisting
         try:
@@ -2681,7 +2735,7 @@ class ChatService:
         # Persist to database
         deck_dict = current_deck.to_dict()
         session_manager = get_session_manager()
-        session_manager.save_slide_deck(
+        save_result = session_manager.save_slide_deck(
             session_id=session_id,
             title=current_deck.title,
             html_content=current_deck.knit(),
@@ -2690,6 +2744,7 @@ class ChatService:
             deck_dict=deck_dict,
             expected_version=expected_version,
         )
+        self._record_deck_version(session_id, save_result)
 
         # Create save point
         try:
@@ -2744,7 +2799,7 @@ class ChatService:
         # Persist to database
         deck_dict = current_deck.to_dict()
         session_manager = get_session_manager()
-        session_manager.save_slide_deck(
+        save_result = session_manager.save_slide_deck(
             session_id=session_id,
             title=current_deck.title,
             html_content=current_deck.knit(),
@@ -2753,6 +2808,7 @@ class ChatService:
             deck_dict=deck_dict,
             expected_version=expected_version,
         )
+        self._record_deck_version(session_id, save_result)
 
         # Create save point
         try:

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/src/api/services/export_job_queue.py

@@ -307,6 +307,9 @@ async def convert_slides_with_progress(
         output_path: Path to save PPTX
         chart_images_per_slide: Chart images per slide
     """
+    from pptx import Presentation
+    from pptx.util import Inches
+
     total_slides = len(slides_html)
 
     # -- Phase 1: Prepare assets + parallel LLM code generation ---------------
@@ -333,17 +336,27 @@ async def convert_slides_with_progress(
         slide_inputs, on_codegen_progress=_on_codegen_progress,
     )
 
-    # -- Phase 2: Sequential execution INSIDE the subprocess jail -------------
-    def _relay(current: int, total: int, message: str) -> None:
-        update_export_progress(job_id, current, total, status_message=message)
+    # -- Phase 2: Sequential execution ----------------------------------------
+    prs = Presentation()
+    prs.slide_width = Inches(10)
+    prs.slide_height = Inches(7.5)
 
-    # to_thread: keep the long blocking jail wait off the event loop (this
-    # coroutine runs on the loop; _relay is fired from the jail's stdout
-    # reader thread and update_export_progress opens its own session per call).
-    await asyncio.to_thread(
-        converter._run_pptx_conversion,
-        codes, slide_inputs, output_path, progress_cb=_relay,
-    )
+    for i, (code, (html_str, _chart_files, _content_files, assets_dir)) in enumerate(
+        zip(codes, slide_inputs), 1,
+    ):
+        update_export_progress(
+            job_id, i, total_slides,
+            status_message=f"Building slide {i}/{total_slides}…",
+        )
+
+        converter._execute_slide(code, prs, html_str, assets_dir, i)
+
+        logger.debug(
+            f"Processed slide {i}/{total_slides}",
+            extra={"job_id": job_id, "slide": i, "total": total_slides},
+        )
+
+    prs.save(output_path)
 
 
 async def export_worker() -> None:

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/src/api/services/session_manager.py

@@ -723,6 +723,33 @@ class SessionManager:
                 "version": deck.version,
             }
 
+    def get_slide_deck_version(self, session_id: str) -> Optional[int]:
+        """Return only the deck's current version number.
+
+        Selects a single column — no deck_json parse, verification merge, or
+        display-name resolution — so it is cheap enough to run on every
+        deck-cache hit (multi-worker cache validation) and before LLM calls
+        (optimistic locking).
+
+        Args:
+            session_id: Session to check (contributor sessions resolve to the
+                owner's deck)
+
+        Returns:
+            Current version number, or None if the session/deck doesn't exist
+        """
+        with get_db_session() as db:
+            try:
+                session = self._get_session_or_raise(db, session_id)
+            except SessionNotFoundError:
+                return None
+            deck_owner = self._get_deck_owner_session(db, session)
+            return (
+                db.query(SessionSlideDeck.version)
+                .filter(SessionSlideDeck.session_id == deck_owner.id)
+                .scalar()
+            )
+
     def get_slide_deck(self, session_id: str) -> Optional[Dict[str, Any]]:
         """Get slide deck for a session with verification merged by content hash.
 

{databricks_tellr_app-0.3.13.dev3 → databricks_tellr_app-0.3.13.dev11}/src/services/google_slides_prompts_defaults.py

@@ -69,9 +69,8 @@ TWO-COLUMN LAYOUT (.two-column):
 - Left: left=0.4", width=4.3"; Right: left=5.0", width=4.5"; Gap: 0.3"
 
 TABLES (<table>):
-- createTable first (rows, cols, size, position), with an explicit objectId
-- AFTER it in the returned list: insertText per cell + updateTextStyle + updateTableCellProperties
-  (the host executes requests in order, so the table exists before its cells are filled)
+- createTable in FIRST batchUpdate (rows, cols, size, position)
+- Then insertText per cell + updateTextStyle + updateTableCellProperties in SECOND batchUpdate
 - updateTableCellProperties requires: 'tableRange': {'location': {'rowIndex': r, 'columnIndex': c}, 'rowSpan': 1, 'columnSpan': 1}
 - Headers: background from CSS (typically dark), text white, font 8pt bold
 - Cells: font 7-8pt, preserve colors, left-align text, right-align numbers
@@ -108,22 +107,17 @@ REGULAR SLIDES:
   Subtitle must fit in 0.28" (10pt ≈ 0.14" per line → 2 lines max).
 
 CHARTS (CRITICAL):
-- If chart images are listed in the user message, you MUST reference them by
-  placeholder — do NOT upload anything, do NOT call any Google API.
-- For each chart image `filename` in assets_dir, add a createImage request with
-  url set to the placeholder 'tellr-asset://' + filename. The host uploads the
-  file to Drive and substitutes the real URL before execution.
-  Example: {'createImage': {'objectId': 'img_' + str(uuid.uuid4())[:8],
-    'url': 'tellr-asset://' + filename,
-    'elementProperties': {'pageObjectId': page_id, 'size': {...}, 'transform': {...}}}}
+- If chart images are listed in the user message, you MUST use them as pre-captured screenshots.
 - Do NOT recreate charts with matplotlib, plotly, or any library. The images are already rendered.
+- Upload to Drive via MediaFileUpload, set anyone/reader permission, then createImage.
 - Chart only (no metric cards): left=0.4", top=1.1", width=9.0", height=4.0" — fills the body zone.
 - Chart + metrics: use CHART + METRICS SIDE-BY-SIDE layout above.
 - If NO chart images are listed, skip chart rendering entirely — do not attempt to recreate.
 
 CONTENT IMAGES (logos, icons — CRITICAL):
-- Reference each content image the same way: url = 'tellr-asset://' + filename.
-- NEVER upload; NEVER create Drive files or set permissions yourself. The host does that.
+- If content images are listed in the user message, you MUST upload and add them to the slide.
+- These are pre-extracted image files (PNG) saved in assets_dir — ready to upload.
+- Upload to Drive the same way as charts: MediaFileUpload → files().create → permissions().create → createImage.
 - Position to match the HTML layout (e.g. a logo in the top-right header area → right-aligned near the title).
 - Typical logo sizes: width=0.6"-1.0", height=0.3"-0.5". Do NOT make logos larger than needed.
 - Place logos/icons in the header zone (top=0.15"-0.4") so they don't overlap body content.
@@ -171,27 +165,23 @@ API PATTERNS (CRITICAL — wrong nesting causes 400 errors):
 
 ALLOWED IMPORTS:
 import os, json, uuid, re
-(Do NOT import googleapiclient — you never touch the network.)
+from googleapiclient.http import MediaFileUpload  # ONLY when uploading images
 
 EXECUTION:
 - ALL code MUST be inside the function body. Nothing at module level except imports and helpers.
-- Build ONE requests list and RETURN it. Never call batchUpdate or any API
-  yourself — the host executes the returned requests in order.
-- Do NOT wrap request-building in try/except.
+- Build requests list, then ONE batchUpdate (except tables need two).
+- Do NOT wrap batchUpdate in try/except.
 - Every createShape needs a unique objectId: 'txt_' + str(uuid.uuid4())[:8]
 
 Return ONLY Python code, no markdown fences."""
 
 # ── Multi-slide prompts ──────────────────────────────────────────────
 
-DEFAULT_GSLIDES_SYSTEM_PROMPT = """Generate Python code with a build_slide_requests(html_str, assets_dir, page_id) function that RETURNS a list of Google Slides batchUpdate request dicts.
+DEFAULT_GSLIDES_SYSTEM_PROMPT = """Generate Python code with add_slide_to_presentation(slides_service, drive_service, presentation_id, page_id, html_str, assets_dir) function.
 
 CONTEXT:
-- You do NOT have and do NOT need any Google service objects or network access.
-- Do NOT call batchUpdate or any other Google API method — never execute requests, upload files, or create permissions yourself.
-- The target slide page already exists; reference it via the page_id argument.
-- Return (do not execute) a Python list of request dicts. The trusted host
-  executes them, uploads any referenced images, and handles chunking/retries.
+- All arguments are pre-created. Do NOT import googleapiclient or build services.
+- The slide page already exists (page_id). Do NOT call createSlide — only add content.
 
 HELPERS (include in your code):
 def emu(inches): return int(inches * 914400)
@@ -202,22 +192,22 @@ def hex_to_rgb(h):
 SLIDE: 16:9 widescreen — 10" × 5.625" (9144000 × 5143500 EMU).
 Bounds: left ≥ 0.4", top ≥ 0.2", left + width ≤ 9.6", top + height ≤ 5.4"
 
-ELEMENT CREATION (append each request dict to a list, then `return requests`):
-- TEXT_BOX: createShape + insertText + updateTextStyle (textRange {'type':'ALL'}) + updateParagraphStyle
+ELEMENT CREATION:
+- TEXT_BOX: createShape + insertText + updateTextStyle (textRange: {'type':'ALL'}) + updateParagraphStyle
 - RECTANGLE/ROUND_RECTANGLE: createShape + updateShapeProperties
 - Size: {'width': {'magnitude': emu(w), 'unit': 'EMU'}, 'height': {'magnitude': emu(h), 'unit': 'EMU'}}
 - Position: {'scaleX':1,'scaleY':1,'translateX':emu(left),'translateY':emu(top),'unit':'EMU'}
 
-The code MUST define build_slide_requests(html_str, assets_dir, page_id) and RETURN the requests list.
+The code MUST define the function add_slide_to_presentation.
 """ + _GSLIDES_SHARED_RULES
 
-DEFAULT_GSLIDES_USER_PROMPT = """Convert the HTML below to a list of Google Slides batchUpdate requests.
+DEFAULT_GSLIDES_USER_PROMPT = """Convert the HTML below to a Google Slide.
 
 {html_content}
 
 {screenshot_note}
 
-Return Python code defining build_slide_requests(html_str, assets_dir, page_id) that returns the requests list."""
+Return Python code with add_slide_to_presentation(slides_service, drive_service, presentation_id, page_id, html_str, assets_dir)."""
 
 # ── Single-slide prompts ─────────────────────────────────────────────