oneshott primpt change

.gitignore

@@ -7,3 +7,4 @@ test*
 all-MiniLM-L6-v2
 cross-encoder/ms-marco-MiniLM-L-6-v2
 test
+RAG/rag_embeddings/[a-z]*

LLM/one_shotter.py

@@ -414,7 +414,8 @@ CRITICAL INSTRUCTIONS:
 4. BE THOROUGH: Don't just use the original context - actively look for and incorporate information from scraped websites
 5. DETAILED EXPLANATIONS: Provide comprehensive, well-structured answers with specific details
 6. IF MISSING INFO: Only state information is missing if it's truly not available in ANY part of the provided context
-
+7. First give the correct answer and then explain in short, you don't need to outline your thought process.
+8. Never make any assumptions on your own.
 The context may contain multiple sections:
 - Original context
 - Additional Information from relevant links  
@@ -425,12 +426,12 @@ USE ALL OF THESE SECTIONS TO PROVIDE COMPLETE ANSWERS.
 Respond in this EXACT JSON format:
 {{
     "answers": [
-        "<Correct Answer to the question 1, with detailed explaination.>",
-        "<Correct Answer to the question 2, with detailed explaination only if question 2 exists.>",
+        "<Correct Answer to the question 1, followed by  explaination.>",
+        "<Correct Answer to the question 2, followed by  explaination only if question 2 exists.>",
         ...
     ]
 }}
-             """)
+        """)
         ])
         
         questions_text = "\n".join([f"{i+1}. {q.strip()}" for i, q in enumerate(questions)])

RAG/rag_modules/query_expansion.py

@@ -42,9 +42,11 @@ Examples of good breakdown:
 
 
 Provide only {QUERY_EXPANSION_COUNT} focused sub-questions, one per line, without numbering or additional formatting:
-
-#### RESPONSE:
-Here are the focused sub queries:
+Example Reponse:
+Here are the focused sub queries
+subquery1
+subquery2 (if exists)
+...
 
 """
 
@@ -60,14 +62,19 @@ Here are the focused sub queries:
             if response:
                 sub_queries = response.strip().split('\n')
                 for query in sub_queries:
-                    if len(expanded_queries) >= QUERY_EXPANSION_COUNT:  # Stop when we have enough
+                    if len(expanded_queries) >= QUERY_EXPANSION_COUNT + 1:  # Stop when we have enough
                         break
                     query = query.strip()
                     # Remove any numbering or bullet points that might be added
                     query = re.sub(r'^[\d\.\-\*\s]+', '', query).strip()
                     if query and len(query) > 10:
                         expanded_queries.append(query)
-            
+
+
+            if len(expanded_queries) > 1:
+                expanded_queries.pop(0)
+        
+
             # If we don't have enough sub-queries, fall back to using the original
             if len(expanded_queries) < QUERY_EXPANSION_COUNT:
                 expanded_queries = [original_query]

config/config.py

@@ -32,13 +32,13 @@ GROQ_MODEL_LITE = "llama3-8b-8192"
 BEARER_TOKEN = os.getenv("BEARER_TOKEN")
 
 # Chunking
-CHUNK_SIZE = 400
-CHUNK_OVERLAP = 100
+CHUNK_SIZE = 400 * 4
+CHUNK_OVERLAP = 100 * 4
 
 # Retrieval Settings
-TOP_K = 15
+TOP_K = 9
 SCORE_THRESHOLD = 0.3 
-RERANK_TOP_K =  9 # 9*400 = 3600, < 4000, some tokens reserved for questions  
+RERANK_TOP_K =  7 # 9*400 = 3600, < 4000, some tokens reserved for questions  
 BM25_WEIGHT = 0.3 
 SEMANTIC_WEIGHT = 0.7
 
@@ -49,7 +49,7 @@ ENABLE_QUERY_EXPANSION = True
 RERANKER_MODEL = "cross-encoder/ms-marco-MiniLM-L-6-v2"
 QUERY_EXPANSION_COUNT = 3 
 SCORE_THRESHOLD = 0.3 
-MAX_CONTEXT_LENGTH = 4000
+MAX_CONTEXT_LENGTH = 4000*4
 
 USE_TOTAL_BUDGET_APPROACH = True  
 
@@ -110,8 +110,8 @@ def get_provider_configs():
     configs["openai"] = [{
             "name": sequence[i],
             "api_key": os.getenv(f"OPENAI_API_KEY_{i}"),
-            "model": os.getenv(f"OPENAI_MODEL_{i}", DEFAULT_GEMINI_MODEL)
-        } for i in range(10) if os.getenv(f"GEMINI_API_KEY_{i}", "")
+            "model": os.getenv(f"OPENAI_MODEL_{i}", DEFAULT_OPENAI_MODEL)
+        } for i in range(10) if os.getenv(f"OPENAI_MODEL_{i}", "")
     ]
     
     return configs

preprocessing/preprocessing_modules/modular_preprocessor.py

@@ -87,7 +87,7 @@ class ModularDocumentPreprocessor:
         Process a single document: download, extract, chunk, embed, and store.
         
         Args:
-            document_url: URL of the PDF document
+            document_url: URL of the document
             force_reprocess: If True, reprocess even if already processed
             timeout: Download timeout in seconds (default: 300s/5min)
             
@@ -106,7 +106,7 @@ class ModularDocumentPreprocessor:
         
         temp_file_path = None
         try:
-            # Step 1: Download PDF
+            # Step 1: Download Document
             temp_file_path, ext = await self.file_downloader.download_file(document_url, timeout=timeout)
             
             if temp_file_path == 'not supported':