Init

app/agents/invoice_agent.py

@@ -4,10 +4,10 @@ from langchain_core.output_parsers import PydanticOutputParser
 from langchain.prompts import PromptTemplate
 from ..core.llm import llm
 from ..schemas import InvoiceInfo
+from typing import List
 
 parser = PydanticOutputParser(pydantic_object=InvoiceInfo)
 
-# The prompt now includes the detailed rules for each field using snake_case.
 invoice_template = """
 You are an expert data entry clerk AI. Your primary goal is to extract information from an invoice image with the highest possible accuracy.
 The document's primary language is '{language}'.
@@ -30,6 +30,9 @@ Carefully analyze the invoice image and extract the following fields according t
 - `customer_address_region`: This is the receiver's region. If not found, find the region of the extracted city or country. If unclear, leave as an empty string.
 - `customer_address_care_of`: This is the receiver's 'care of' (c/o) line. If not found or unclear, leave as an empty string.
 - `billo_id`: To find this, think step-by-step: 1. Find the customer_address. 2. Scan the address for a pattern of three letters, an optional space, three digits, an optional dash, and one alphanumeric character (e.g., 'ABC 123-X' or 'DEF 456Z'). 3. If found, extract it. If not found or unclear, leave as an empty string.
+- `bank_giro`: If found, extract the bank giro number. It often follows patterns like 'ddd-dddd', 'dddd-dddd', or 'dddddddd #41#'. If not found or unclear, leave as an empty string.
+- `plus_giro`: If found, extract the plus giro number. It often follows patterns like 'ddddddd-d #16#', 'ddddddd-d', or 'ddd dd dd-d'. If not found or unclear, leave as an empty string.
+- `customer_ssn`: If found, extract the customer social security number (personnummer). It follows the pattern 'YYYYMMDD-XXXX' or 'YYMMDD-XXXX'. If not found or unclear, leave as an empty string.
 - `line_items`: Extract all line items from the invoice. For each item, extract the `description`, `quantity`, `unit_price`, and `total_price`. If a value is not present, leave it as null.
 
 ## Example:
@@ -41,33 +44,31 @@ If the invoice shows a line item "Consulting Services | 2 hours | $100.00/hr | $
   "unit_price": 100.00,
   "total_price": 200.00
 }}
-
+```
 Your Task:
 Now, analyze the provided image and output the full JSON object according to the format below.
 {format_instructions}
 """
-
 invoice_prompt = PromptTemplate(
     template=invoice_template,
     input_variables=["language"],
-    partial_variables={"format_instructions": parser.get_format_instructions()},
+    partial_variables={"format_instructions": parser.get_format_instructions()}
 )
 
-async def agent_extract_invoice_info(image_base64: str, language: str) -> InvoiceInfo:
-    """Agent 3: Extracts invoice information from an image, aware of the document's language."""
+async def agent_extract_invoice_info(images_base64: List[str], language: str) -> InvoiceInfo:
+    """Agent 3: Extracts invoice information from a list of images, aware of the document's language."""
     print(f"--- [Agent 3] Calling multimodal LLM to extract invoice info (Language: {language})...")
 
     prompt_text = await invoice_prompt.aformat(language=language)
 
-    msg = HumanMessage(
-        content=[
-            {"type": "text", "text": prompt_text},
-            {
-                "type": "image_url",
-                "image_url": f"data:image/png;base64,{image_base64}",
-            },
-        ]
-    )
+    content_parts = [{"type": "text", "text": prompt_text}]
+    for image_base64 in images_base64:
+        content_parts.append({
+            "type": "image_url",
+            "image_url": {"url": f"data:image/png;base64,{image_base64}"},
+        })
+
+    msg = HumanMessage(content=content_parts)
 
     chain = llm | parser
     invoice_info = await chain.ainvoke([msg])