Update app.py

app.py

@@ -1,5 +1,6 @@
 import os
 import io
+import re
 import json
 import zipfile
 import asyncio
@@ -236,6 +237,41 @@ def prepare_file(file_bytes: bytes, content_type: str, filename: str) -> tuple[b
         with open(out_path, "rb") as f:
             return f.read(), Path(filename).stem + ".pdf"
 
+
+# ---------------------------------------------------------------------------
+# Text cleanup — strip base64-encoded image blobs from OCR output
+# ---------------------------------------------------------------------------
+
+def clean_raw_text(text: str) -> str:
+    """
+    Remove base64-encoded image data that Sarvam Document Intelligence
+    carries over from PDFs/images into the markdown output.
+
+    These blobs are never useful for field extraction and can easily push
+    the prompt far beyond sarvam-m's 7 192-token context window.
+
+    Two patterns are handled:
+      1. Markdown image syntax:  ![alt](data:image/png;base64,<data>)
+      2. Bare data URIs that leaked outside markdown syntax
+    """
+    # Pattern 1 — markdown image embed
+    text = re.sub(
+        r'!\[[^\]]*\]\(data:[a-zA-Z]+/[a-zA-Z+\-]+;base64,[A-Za-z0-9+/=\s]+\)',
+        '',
+        text,
+        flags=re.DOTALL,
+    )
+    # Pattern 2 — bare data URI (must be at least 100 chars of base64 to avoid
+    #              accidentally removing short legitimate strings)
+    text = re.sub(
+        r'data:[a-zA-Z]+/[a-zA-Z+\-]+;base64,[A-Za-z0-9+/=\s]{100,}',
+        '',
+        text,
+        flags=re.DOTALL,
+    )
+    return text.strip()
+
+
 # ---------------------------------------------------------------------------
 # Step 1: Extract raw text via Sarvam Document Intelligence SDK
 # ---------------------------------------------------------------------------
@@ -322,7 +358,6 @@ Respond with the JSON object only."""
 
     raw = resp.json()["choices"][0]["message"]["content"].strip()
 
-    import re
     raw = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL).strip()
 
     if raw.startswith("```"):
@@ -410,6 +445,9 @@ async def extract_document(
         )
         raise
 
+    # Strip base64 image blobs before sending to sarvam-m
+    raw_text = clean_raw_text(raw_text)
+
     try:
         extracted = await extract_json_via_chat(document_type, raw_text, param_list)
     except HTTPException as exc: