Init project

app/agents/invoice_agent.py

@@ -0,0 +1,74 @@
+# app/agents/invoice_agent.py
+from langchain_core.messages import HumanMessage
+from langchain_core.output_parsers import PydanticOutputParser
+from langchain.prompts import PromptTemplate
+from ..core.llm import llm
+from ..schemas import InvoiceInfo
+
+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}'.
+
+## Instructions:
+Carefully analyze the invoice image and extract the following fields according to these specific rules. Do not invent information. If a field is not found or is unclear, follow the specific instruction for that field.
+
+- `date`: Extract in YYYY-MM-DD format. If unclear, leave as an empty string.
+- `invoice_number`: If not found or unclear, leave as an empty string.
+- `supplier_number`: This is the organisation number. If not found or unclear, leave as an empty string.
+- `biller_name`: This is the sender's name. If not found or unclear, leave as an empty string.
+- `amount`: Extract the final total amount and format it to a decimal number. If not present, leave as null.
+- `customer_name`: This is the receiver's name. Ensure it is a name and clear any special characters. If not found or unclear, leave as an empty string.
+- `customer_address`: This is the receiver's full address. Put it in one line. If not found or unclear, leave as an empty string.
+- `customer_address_line`: This is only the street address line from the receiver's address. If not found or unclear, leave as an empty string.
+- `customer_address_city`: This is the receiver's city. If not found, try to find any city in the document. If unclear, leave as an empty string.
+- `customer_address_country`: This is the receiver's country. If not found, find the country of the extracted city. If unclear, leave as an empty string.
+- `customer_address_postal_code`: This is the receiver's postal code. If not found or unclear, leave as an empty string.
+- `customer_address_apartment`: This is the receiver's apartment or suite number. If not found or unclear, leave as an empty string.
+- `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.
+- `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:
+If the invoice shows a line item "Consulting Services | 2 hours | $100.00/hr | $200.00", the output for that line item should be:
+```json
+{{
+  "description": "Consulting Services",
+  "quantity": 2,
+  "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()},
+)
+
+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."""
+    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}",
+            },
+        ]
+    )
+
+    chain = llm | parser
+    invoice_info = await chain.ainvoke([msg])
+    return invoice_info