fix: pre-extract PDF text in API to share with Celery worker

Railway runs the API and Worker in separate containers with isolated
filesystems. The worker cannot read files saved by the API.

Fix: extract text from the PDF immediately in the upload endpoint and
store it in document.raw_text (column already existed). The pipeline's
step 1 now checks raw_text first — if populated it uses that and builds
a synthetic page_map, skipping the file read entirely. The file-read
path is kept as a fallback for local dev where the filesystem is shared.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: MLaitarovsky

apps/api/app/routers/documents.py

@@ -1,5 +1,6 @@
 """Document endpoints — upload, list, detail, delete."""
 
+import logging
 import os
 import uuid
 
@@ -20,8 +21,10 @@
     DocumentUploadResponse,
 )
 from app.tasks.process_document import process_document
+from app.utils.pdf_parser import extract_text_from_pdf
 
 router = APIRouter(prefix="/api/documents", tags=["documents"])
+logger = logging.getLogger(__name__)
 
 MAX_FILE_SIZE = 10 * 1024 * 1024  # 10 MB
 
@@ -73,13 +76,29 @@ async def upload_document(
     with open(file_path, "wb") as f:
         f.write(contents)
 
+    # Pre-extract text from the PDF here in the API process.
+    # The API and Celery worker run in separate containers on Railway and do
+    # NOT share a filesystem — the worker cannot open a file saved by the API.
+    # By extracting text now and storing it in raw_text, the worker can read
+    # it from the database and skip the file-read step entirely.
+    raw_text: str | None = None
+    page_count: int | None = None
+    try:
+        extracted_text, page_map = extract_text_from_pdf(file_path)
+        raw_text = extracted_text
+        page_count = len(page_map)
+    except Exception as exc:
+        logger.warning("Pre-extraction failed, worker will attempt: %s", exc)
+
     # Create the document record
     document = Document(
         team_id=user.team_id,
         uploaded_by=user.id,
         filename=safe_name,
         file_path=file_path,
         file_size_bytes=len(contents),
+        raw_text=raw_text,
+        page_count=page_count,
         status="uploaded",
     )
     db.add(document)

apps/api/app/services/extraction_pipeline.py

@@ -118,11 +118,32 @@ def run(self) -> None:
 
     def _step1_extract_text(self) -> None:
         self._publish(1, "Extracting text from PDF...", 10)
-        logger.info("Step 1/5 — Extracting text from %s", self.document.file_path)
 
+        if self.document.raw_text:
+            # Text was pre-extracted by the API on upload.  In production the
+            # API and Worker run in separate containers that don't share a
+            # filesystem, so we read from the database instead of the file.
+            logger.info(
+                "Step 1/5 — Using pre-extracted text (%d chars)",
+                len(self.document.raw_text),
+            )
+            self.full_text = self.document.raw_text
+            # Build a synthetic page_map so the rest of the pipeline works.
+            # Page numbers on clauses will all be 1, which is acceptable.
+            self.page_map = [
+                {
+                    "page": 1,
+                    "start_char": 0,
+                    "end_char": len(self.full_text),
+                    "text": self.full_text,
+                }
+            ]
+            return
+
+        # Fallback: extract from file directly (local dev with shared filesystem).
+        logger.info("Step 1/5 — Extracting text from %s", self.document.file_path)
         self.full_text, self.page_map = extract_text_from_pdf(self.document.file_path)
 
-        # Update the document with extracted text and page count
         self.document.raw_text = self.full_text
         self.document.page_count = len(self.page_map)
         self.db.commit()