update

app.py

@@ -52,49 +52,35 @@ llm_vision = ChatOpenAI(
 )
 
 # Define the prompt for categorization
-text_prompt = ChatPromptTemplate.from_template("""
-You are an expert at analyzing receipt text extracted with OCR. 
-The text may contain noise or extra characters, but focus only on useful parts.
+UNIFIED_PROMPT_RULES = """
+You are an expert at analyzing receipt content to extract structured fields.
 
 Rules:
-1. **Venue**:
-   - This is the store/restaurant/tuck shop/retailer where the receipt comes from.
-   - It is usually near the top of the receipt (sometimes next to a logo or under "Shop/Store/Restaurant").
-   - If multiple candidates appear, pick the most likely actual business name (ignore words like "Customer Copy", "Approved", "MasterCard").
-   - If both a brand name and location/branch are present, include both (e.g. "Checkers - Brooklyn Mall").
-   - Ignore banks, payment processors, or card brands (e.g., "Nedbank", "FNB", "MasterCard", "Visa") unless they are the ONLY candidate.
-   - If uncertain, check for context clues like items, meals, or store names to identify the real merchant.
-   - If you are not confident, return `null`.
-
-2. **Date**:
-   - Look for patterns like `DD-MM-YY`, `YYYY-MM-DD`, `DD/MM/YYYY`, or combined with time (e.g., `2025-09-15 14:23`).
-   - If only time is found without a date, return `null`.
-
-3. **Total**:
-   - Find the final purchase amount.
-   - Prioritize lines with "TOTAL", "Amount Due", "Purchase", "Grand Total".
-   - Include the **currency symbol/code** if present (e.g., `R238.00`, `$12.50`).
-   - If no currency is visible, return just the number.
-
-4. **Category**:
-   - Based on the venue, infer the main category: groceries, restaurant, fuel, retail, pharmacy, etc.
-   - If the venue is unclear, leave category as `null`.
-
-5. **Output format**:
-   Return strictly valid JSON with these keys:
-   {{
-     "venue": "...",
-     "date": "...",
-     "total": "...",
-     "category": "..."
-   }}
-   - Do not add extra text or explanation outside the JSON.
-
----
-
-Receipt OCR text:
-{ocr_text}
-""")
+1. Venue:
+   - The store/restaurant/retailer name (often at the top). If brand + branch appear, include both.
+   - Ignore banks, payment processors, and card brands unless they are the only candidate.
+   - If uncertain, return null.
+
+2. Date:
+   - Accept formats like DD-MM-YY, YYYY-MM-DD, DD/MM/YYYY, optionally with time.
+   - If only a time is present (no date), return null.
+
+3. Total:
+   - The final purchase amount; prioritize lines with TOTAL/Amount Due/Purchase/Grand Total.
+   - Include currency symbol/code if shown (e.g., R238.00, $12.50). If not visible, return just the number.
+
+4. Category:
+   - Infer from venue/items if possible (groceries, restaurant, fuel, retail, pharmacy, etc.).
+   - If unclear, return null.
+
+5. Output format:
+   - Return strictly valid JSON with keys: {"venue", "date", "total", "category"} and no extra commentary.
+"""
+
+text_prompt = ChatPromptTemplate.from_template(
+    UNIFIED_PROMPT_RULES
+    + "\n---\n\nReceipt OCR text:\n{ocr_text}\n"
+)
 
 
 def run_langchain_with_text(receipt_text: str, image_path: str):
@@ -202,15 +188,7 @@ async def process_receipt_vision(file: UploadFile = File(...)):
 
     data_url = encode_image_to_data_url(cropped, format="JPEG")
 
-    vision_instructions = (
-        "You are an expert at reading receipts from an image. "
-        "Extract these fields as strict JSON: {\\n"
-        "  \"venue\": string or null,\\n"
-        "  \"date\": string or null,\\n"
-        "  \"total\": string or number, include currency if shown,\\n"
-        "  \"category\": string or null\\n"
-        "}. Do not include any extra commentary."
-    )
+    vision_instructions = UNIFIED_PROMPT_RULES
 
     message = HumanMessage(
         content=[
@@ -223,9 +201,14 @@ async def process_receipt_vision(file: UploadFile = File(...)):
 
     return {
         "status": "success",
-        "cropped_file": save_path,
-        "bbox": bbox,
-        "extracted": getattr(result, "content", str(result)),
+        "receipts": [
+            {
+                "cropped_file": save_path,
+                "bbox": bbox,
+                "ocr_text": None,
+                "extracted": getattr(result, "content", str(result)),
+            }
+        ],
     }
 
 if __name__ == "__main__":