Sync from GitHub: c92cec18d31886d94642ced850e1385758cfff7d

inference.py

@@ -67,34 +67,29 @@ Output rules:
 COMBINED_REASONING_EXTRACTION_PROMPT = """
 Analyze this Indian tractor invoice using Chain of Thought reasoning.
 
-First, share your observations about the 4 key fields:
-
-DEALER NAME:
-- Where do you see it? (header, letterhead, stamp)
-- What language? What exact text?
+First, share your observations about the 2 key fields:
 
 MODEL NAME:
-- How is it presented? (checkbox/handwritten/printed)
-- If checkboxes, which is marked?
+- How is it presented? (checkbox/handwritten/printed or any other way)
+- If a checkboxes or anything marked?
 - What exact text do you see?
+- Based on this, which model name you think is involved in the deal.
 
 HORSE POWER:
 - Where is HP mentioned?
 - Explicit text like "49 HP" or in checkbox?
-- Which value is selected?
+- Which value is marked?
 - HP must come from explicit HP text only, never from model numbers
+- If only one value for HP is associated with the correct Model name, it is the models HP.
+- If multiple HP values are associated with the correct Model, the correct HP will be usuall marked.
 
-ASSET COST:
-- Where is the final total?
-- Which amount is after all taxes?
-- What exact amount with currency?
 
 After reasoning, extract the fields.
 
 Return ONLY valid JSON:
 {{
   "reasoning": "your observations and thoughts here",
-  "dealer_name": string,
+  "dealer_name": "string",
   "model_name": string,
   "horse_power": number,
   "asset_cost": number
@@ -108,6 +103,16 @@ Rules for extraction:
 - Dealer is usually at top header
 - If handwriting unclear, make best interpretation but preserve language
 
+
+Extraction hints:
+- Dealer name is usually at the top header or company name.
+- Model name often appears near words like Model, Tractor, Variant.
+- Horse power must come ONLY from explicit HP text, never from model numbers.
+- Horse power may appear as "HP", handwritten like "49 HP", "63hp", "HP-30".
+- Remove commas and currency symbols from numbers only.
+- If handwriting is unclear, make your best reasonable interpretation of the characters — but preserve language.
+
+
 Output ONLY valid JSON, no markdown.
 """