changes

agents/agent2_tenancy_faq.py

@@ -4,7 +4,6 @@ from geotext import GeoText
 import spacy
 import spacy.cli
 
-# Initialize LLaMA and spaCy
 llm = LLaMAHelper()
 
 try:
@@ -60,6 +59,5 @@ def handle_tenancy_query(user_query, user_context, history=[], location_method="
         prompt += f"\n\nPrevious conversation:\n{chat_context}"
 
     prompt += f"\n\nUser's current question: {user_query}\n\nGive a concise and helpful answer. If needed, ask a follow-up question to clarify."
-    print(f'prompt for tenacy faq is {prompt}')
-    reply = llm.chat(system_prompt, prompt, temperature=0.7)
+    reply = llm.chat(system_prompt, prompt, temperature=0.6)
     return reply

app.py

@@ -5,6 +5,13 @@ from agents.agent2_tenancy_faq import handle_tenancy_query
 from PIL import Image
 import torch
 import hashlib
+import subprocess
+import sys
+import os
+
+# Install requirements if needed
+requirements_file = os.path.join(os.path.dirname(__file__), "requirements.txt")
+subprocess.check_call([sys.executable, "-m", "pip", "install", "-r", requirements_file])
 
 # Helper to generate MD5 hash from image
 def get_image_hash(image):

utils/captioning.py

@@ -1,7 +1,6 @@
 from transformers import BlipProcessor, BlipForConditionalGeneration, AutoProcessor, AutoModelForCausalLM
 from PIL import Image
 import torch
-from ultralytics import YOLO  # You need to install: pip install ultralytics
 from transformers import CLIPProcessor, CLIPModel
 
 class ImageCaptioning:
@@ -19,12 +18,8 @@ class ImageCaptioning:
         self.clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
         self.clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(self.device)
 
-        # # Load YOLO
-        # self.yolo_model = YOLO("yolov8n.pt")  # You can use yolov8s.pt or others
-
     def generate_caption_blip(self, image):
         inputs = self.blip_processor(images=image, return_tensors="pt").to(self.device)
-        print(f"Inputs keys: {inputs.keys()}")
         with torch.no_grad():
             output = self.blip_model.generate(**inputs)
         caption = self.blip_processor.decode(output[0], skip_special_tokens=True)
@@ -32,30 +27,10 @@ class ImageCaptioning:
     
     def generate_caption_git(self, image):
         inputs = self.git_processor(images=image, return_tensors="pt").to(self.device)
-        print(f"Inputs keys: {inputs.keys()}")
         with torch.no_grad():
             generated_ids = self.git_model.generate(**inputs)
         caption = self.git_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
         return caption, self.compute_logprob(self.git_model, inputs, generated_ids, self.git_processor)
-
-    # def generate_caption_yolo(self, image):
-    #     # Run YOLO detection
-    #     results = self.yolo_model(image)
-    #     detections = results[0].boxes.data  # [x1, y1, x2, y2, conf, class]
-    #     names = results[0].names
-
-    #     if len(detections) == 0:
-    #         return "No objects detected", 0.0
-
-    #     # Get top class labels with confidence
-    #     label_conf_pairs = [(names[int(cls)], float(conf)) for *_, conf, cls in detections]
-    #     label_conf_pairs.sort(key=lambda x: x[1], reverse=True)
-
-    #     top_labels = list({label for label, _ in label_conf_pairs[:5]})  # top 5 unique labels
-    #     avg_conf = sum([conf for _, conf in label_conf_pairs[:5]]) / len(top_labels)
-
-    #     caption = "Image contains: " + ", ".join(top_labels)
-    #     return caption, avg_conf
     
     def generate_caption_clip(self, image):
         # Step 1: Generate caption candidates
@@ -97,11 +72,8 @@ class ImageCaptioning:
 
         return -outputs.loss.item()  # Higher is better
 
-
-
     def get_best_caption(self, image):
     # This runs BLIP and GIT, then scores both with CLIP to pick the best caption
         caption, score = self.generate_caption_clip(image)
-        print(f"Selected Caption: {caption} | Confidence: {score}")
         return caption, score
 

utils/llm_utils.py

@@ -22,13 +22,8 @@ class LLaMAHelper:
             tokenizer=self.tokenizer,
             device=0 if torch.cuda.is_available() else -1
         )
-
-    #     self.text_classifier = pipeline("zero-shot-classification", model="MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli")
-        
-    # def classifier(self, text, candidate_labels):
-    #     return self.text_classifier(text, candidate_labels)
     
-    def chat(self, system_prompt, prompt, max_new_tokens=300, temperature=0.5):
+    def chat(self, system_prompt, prompt, max_new_tokens=1200, temperature=0.5):
         messages = [
                     {"role": "system", "content": system_prompt},
                     {"role": "user", "content": prompt},
@@ -39,7 +34,5 @@ class LLaMAHelper:
                 full_response = outputs[0]["generated_text"][-1]["content"].lower()
             else:
                 full_response = outputs[0]["generated_text"][-1].lower()
-
-        print('response from LLM is', full_response)
         return full_response.replace(prompt, "").strip()
 

utils/routing.py

@@ -1,28 +1,22 @@
 # agent_router.py
 import re
 
-# Router to manage multi-agent classification
-# agent_router.py
-import re
 
 def route_agent(text, has_image):
     """
     Determines which agent should handle the query based on image presence and content type.
     """
     if has_image:
-        return "agent1"  # Image-based input => Image issue agent
+        return "agent1" 
 
-    # Check for tenancy-related keywords
     tenancy_keywords = [
         "rent", "lease", "tenant", "landlord", "agreement", "deposit",
         "eviction", "notice", "contract", "housing law", "tenancy", "sublet"
     ]
 
-    # Lowercase and look for keywords
     if any(word in text.lower() for word in tenancy_keywords):
         return "agent2"
 
-    # Fallback — you could route this to a clarification step instead
     return "agent2"
 
 def clarify_prompt():