fine tuned

attr_extraction/services.py

@@ -617,65 +617,196 @@ Do not identify or include any additional attributes. The 'additional' section m
             output_example_additional = '  "additional": {}'
 
         prompt = f"""
-You are a product-attribute classifier and validator.
-Understand the product text very deeply. If the same product is available somewhere online, use that knowledge to predict accurate attribute values.
-Do not depend only on word-by-word matching from the product text - interpret the meaning and suggest attributes intelligently.
-Pick the *closest meaning* value from the allowed list, even if not an exact word match.
-I want values for all mandatory attributes.
-If a value is not found anywhere, the source should be "inferred".
-Note: Source means from where you have concluded the result. Choose one of these value <{source_hint}>
-Do not give "Condition" attribute ever.
-
-
-ALLOWED VALUES (MANDATORY):
-{allowed_text}
-
-Note: "Strictly" return multiple values for these attributes: {multiple_text}. These values must be most possible values from the list and should be max 2 values. 
-
-{user_values_section}
+You are an expert product attribute extraction system. Your task is to classify products into predefined categories with 100% accuracy.
+
+CORE RULES - FOLLOW STRICTLY:
+1. You MUST ONLY use values from the ALLOWED VALUES lists provided below
+2. NEVER invent, create, or infer values that don't exist in the allowed lists
+3. NEVER use synonyms, variations, or similar words - ONLY exact matches from the lists
+4. Each value you return MUST be an EXACT COPY from the allowed list (case-sensitive)
+5. If you cannot find a perfect match, choose the CLOSEST semantic match from the allowed list
+6. When unsure, default to the most generic option from the allowed list
+7. CRITICAL GUARDRAIL: NEVER use a value allowed for one attribute (e.g., T-Shirts from t_shirt_type) as the value for a different attribute (e.g., style_clothing_top). The lists are entirely separate.
+8. NEVER invent, create, or infer values that don't exist in the allowed lists.
+9. Special Instruction for T-Shirts: If the product text describes a generic T-shirt (Crew Neck, V-Neck, etc.) which is a basic shirt worn by pulling it over the head, you MUST use the value Pullover. This is the most appropriate generic style available in this specific list.
+
+
+=====================
+ATTRIBUTE DEFINITIONS + ALLOWED VALUES
+=====================
+
+1️⃣ **style_clothing_top** (Mandatory)
+Defines the overall silhouette or construction of the top.
+
+Allowed values + definitions:
+- Bandeau: Strapless, tight top covering just bust.
+- Blouse: Loose, dressy top, usually with sleeves.
+- Camisole: Thin-strapped sleeveless lightweight top.
+- Chemise: Straight, loose unshaped top.
+- Cocoon: Rounded, oversized body silhouette.
+- Corset: Structured top shaping waist with boning.
+- Crop: Short top exposing midriff above waist.
+- Cutout: Top with intentionally open fabric areas.
+- Duster: Long open-front flowy layer.
+- Flounce: Top featuring decorative ruffles.
+- Full Zip: Zipper opening entire length front.
+- Guide: Top with reference markers or functional guides.
+- Half Zip: Zipper halfway down front.
+- High-Low: Back hem longer than front.
+- Hoodie: Top with a hood attached.
+- Muscle: Sleeveless with wide armholes, athletic.
+- Peasant: Loose boho style with gathered neckline/sleeves.
+- Peplum: Fitted top with flared waist ruffle.
+- Pullover: Worn by pulling over head; no front opening.
+- Quarter Zip: Short zipper from collar ~¼ length down.
+- Raglan: Sleeves extend to collar with diagonal seams.
+- Ringer: Contrast colored sleeve cuffs + neckline band.
+- Rugby: Thick striped or solid collared sport-style top.
+- Smocked: Gathered elastic shirring for stretch texture.
+- Swing: A-line flare from bust downward.
+- Torsette: Corset-like, bust exposed for layering.
+- Tube: Strapless elongated top; longer than bandeau.
+- Zip-Up: Top with zipper closure (partial/full).
+
+
+2️⃣ **shirt_neck_style** (Mandatory)
+Describes the neckline’s shape and construction.
+
+Allowed values + definitions:
+- Boat Neck: Wide neckline shoulder-to-shoulder.
+- Caged Neck: Multiple straps forming cage-like design.
+- Choker Neck: Tight high neck like choker band.
+- Collared: Fold-over collar shirt/polo style.
+- Cowl Neck: Draped soft neckline folds.
+- Crew Neck: Close-fitting round classic neckline.
+- Deep V-Neck: Deep V shape below chest level.
+- Drape Neck: Softly draped neckline less than cowl.
+- Funnel Neck: Short stand-up collar not folded.
+- Halter: Straps around neck leaving shoulders bare.
+- Henley: Round neck with button placket.
+- High Neck: More neck coverage, no fold.
+- Hooded: Neck includes a hood.
+- Jewel Neck: High round neck at base of throat.
+- Keyhole Neck: Slit/hole opening at neckline.
+- Lace Neckline: Lace material used around neckline.
+- Mock Neck: Short raised collar, not folded.
+- Notch Neck: Small V cut in round neckline.
+- Open: General wide/open neckline shape.
+- Plunge: Very deep V/U revealing cleavage.
+- Roll Neck: Loose rolled turtleneck style.
+- Round Neck: Standard circular neckline.
+- Round Neckline: Same as Round Neck.
+- Scoop Neck: Wide deep U-shaped neckline.
+- Scrunch Neck: Gathered/scrunched fabric at neckline.
+- Slit Neck: Small vertical slit opening at front.
+- Square Neckline: Straight edged square neck.
+- Sweetheart: Heart-shaped neckline contour.
+- Tie Neck: Ties or bow at neckline.
+- Turtleneck: High folded collar covering full neck.
+- V-Neck: V-shaped neckline (not deep).
+- Wide Neck: Broad neckline toward shoulders.
+- Zip Mock Neck: Mock neck with zipper.
+
+3️⃣ **t_shirt_type** (Mandatory)
+Defines the category/design purpose of the T-shirt.
+
+Allowed values + definitions:
+- Babydoll T-Shirt: Feminine fitted shirt with flared hem.
+- Classic T-Shirt: Standard basic crew tee.
+- Graphic Tees: T-shirts featuring printed graphics.
+- Pocket Tee: T-shirt having a chest pocket.
+- T-Shirts: General type when no specific style is clear.
+- Tank Tops: Sleeveless shirts with shoulder straps.
+
+=====================
+MANDATORY ATTRIBUTE RULES
+=====================
+
+MULTI-VALUE ATTRIBUTES (Return exactly 2 most relevant values ONLY):
+{multiple_text}
+
+SINGLE-VALUE ATTRIBUTES (Return exactly 1 value for all other attributes)
+
+=====================
+INPUT PRODUCT DETAILS
+=====================
+PRODUCT INFORMATION TO ANALYZE:
+{product_text}
 
 {additional_instructions}
 
-PRODUCT TEXT:
-{product_text}
+=====================
+EXTRACTION STRATEGY
+=====================
+Step 1: Read product text carefully
+Step 2: Identify strong style/type/neck indicators
+Step 3: Choose closest exact match from allowed list
+Step 4: VALIDATE chosen values exist in allowed list
+
+=====================
+FORBIDDEN ACTIONS
+=====================
+:x: NEVER invent values
+:x: NEVER modify spelling/capitalization
+:x: NEVER use non-allowed synonyms
+:x: NEVER return attributes outside allowed lists
+:x: NEVER use descriptive words like “short sleeve”, “women top”, “graphic print” unless EXACT in list
+
+=====================
+OUTPUT FORMAT
+=====================
+Return ONLY this JSON structure — no markdown, no explanation:
 
-OUTPUT (strict JSON only):
 {{
   "mandatory": {{
-    "<attr>": [{{
-      "value": "<chosen_value>", 
-      "source": "<{source_hint}>",
-      "reason": "Explanation of why this value was chosen. If user provided a value, explain why you agreed/disagreed with it.",
-      "original_value": "<user_entered_value_if_provided>",
-      "decision": "accepted|rejected|not_provided"
+    "style_clothing_top": [{{
+      "value": "<exact value from allowed list>",
+      "source": "{source_hint}",
+      "reason": "Explain specific mapping from product text"
+    }}],
+    "shirt_neck_style": [{{
+      "value": "<exact value from allowed list>",
+      "source": "{source_hint}",
+      "reason": "Explain specific mapping from product text"
+    }}],
+    "t_shirt_type": [{{
+      "value": "<exact value from allowed list>",
+      "source": "{source_hint}",
+      "reason": "Explain specific mapping from product text"
     }}]
   }},
 {output_example_additional}
 }}
 
-RULES:
-- For each mandatory attribute with a user-entered value, include "original_value" and "decision" fields
-- "decision" values: "accepted" (used user's value), "rejected" (used different value), "not_provided" (no user value given)
-- "reason" must explain your choice, especially when rejecting user input
-- For 'multiple' attributes, always give multiple values for those attributes, choose wisely and max 2 values per attribute that are very close. 
-- Source must be one of: {source_hint}
-- Be honest and specific in your reasoning.
-- Return ONLY valid JSON
+=====================
+FINAL VALIDATION BEFORE RESPONDING
+=====================
+✓ All values EXACT from allowed list
+✓ Multi-value attributes have exactly 2 values
+✓ Single-value attributes have exactly 1 value
+✓ “source” must be one of: {source_hint}
+✓ Reasons clearly explain mapping
+✓ Pure JSON — no markdown wrapper
+
 """
 
+
+        print(f"Prompt to the llm is: {prompt}")
+
         payload = {
             "model": model,
             "messages": [
                 {"role": "system", "content": "You are a JSON-only extractor and validator. Always provide clear reasoning for your decisions."},
                 {"role": "user", "content": prompt},
             ],
-            "temperature": 0.3,
+            "temperature": 0.0,
             "max_tokens": 2000,  # Increased for reasoning
         }
 
         try:
             raw = ProductAttributeService._call_llm(payload)
             logger.info("Raw LLM response received")
+            print(raw)
             cleaned = ProductAttributeService._clean_json(raw)
             parsed = json.loads(cleaned)
         except Exception as exc:

attr_extraction/urls.py

@@ -20,7 +20,7 @@ from .views import CacheManagementView, CacheStatsView
 urlpatterns = [
     # Existing endpoints
     path('extract/', ExtractProductAttributesView.as_view(), name='extract-attributes'),
-    path('batch-extract-actual/', BatchExtractProductAttributesView.as_view(), name='batch-extract-attributes'),
+    path('batch-extract/', BatchExtractProductAttributesView.as_view(), name='batch-extract-attributes'),
     path('products/', ProductListView.as_view(), name='product-list'),
     path('products/upload-excel/', ProductUploadExcelView.as_view(), name='product-upload-excel'),
     path('products/upload-attributes/', ProductAttributesUploadView.as_view(), name='product-upload-attributes'),
@@ -38,7 +38,7 @@ urlpatterns = [
     path('cache/management/', CacheManagementView.as_view(), name='cache-management'),
     path('cache/stats/', CacheStatsView.as_view(), name='cache-stats'),
 
-    path('batch-extract/', ReadLocalJSONView.as_view(), name='ReadLocalJSONView'),
+    path('batch-extract-json/', ReadLocalJSONView.as_view(), name='ReadLocalJSONView'),
 
 ]
 

attr_extraction/views.py

@@ -566,7 +566,8 @@ class BatchExtractProductAttributesView(APIView):
                     extract_additional=extract_additional,
                     multiple=multiple,
                     use_cache=True,
-                    user_entered_values=user_entered_values  # 🆕 PASS USER VALUES
+                    user_entered_values=""  # 🆕 PASS NULL
+                    # user_entered_values=user_entered_values  # 🆕 PASS USER VALUES
                 )
 
                 # NOTE: Original values are now part of LLM response with reasoning