relevant update

service/rag_service.py

@@ -1,62 +1,63 @@
-import pickle
-from pathlib import Path
-
 from service.data_loader_service import CSVDataLoader
 from service.embedded_service import EmbeddingService
 from service.vector_store_service import VectorStoreService
 from service.llm_service import LLMService
+from pathlib import Path
+import pickle
 
-# -----------------------------
-# Constants / Cache
-# -----------------------------
-CACHE_PATH = Path("embeddings.pkl")
-CSV_FILE = "final_data_set(in).csv"  # your QA CSV file
+# -------------------------------
+# Setup
+# -------------------------------
 
-# -----------------------------
-# Initialize services
-# -----------------------------
+CACHE = Path("embeddings.pkl")
+
+# Embedder and LLM
 embedder = EmbeddingService()
 llm = LLMService()
 
 # Load documents
-loader = CSVDataLoader(CSV_FILE)
-documents = loader.load_qa_pairs()  # ["Q: ...\nA: ...", ...]
+loader = CSVDataLoader("final_data_set(in).csv")
+documents = loader.load_qa_pairs()
 
 # Load or compute embeddings
-if CACHE_PATH.exists():
-    with CACHE_PATH.open("rb") as f:
+if CACHE.exists():
+    with CACHE.open("rb") as f:
         embeddings = pickle.load(f)
 else:
     embeddings = embedder.embed(documents)
-    with CACHE_PATH.open("wb") as f:
+    with CACHE.open("wb") as f:
         pickle.dump(embeddings, f)
 
 vector_store = VectorStoreService(embeddings, documents)
 
+# -------------------------------
+# Generate answer function
+# -------------------------------
 
-# -----------------------------
-# Generate Answer Function
-# -----------------------------
-def generate_answer(question: str, k: int = 3) -> str:
+def generate_answer(question: str, k: int = 3, min_similarity: float = 0.65) -> str:
     """
-    Generate an answer strictly from context.
-    If the answer is not present, return the context instead of making up a solution.
+    Generates answer using RAG (retrieval + LLM).
+    Only includes context with similarity above threshold.
+    Returns fallback if no relevant context.
     """
 
+    # Compute embedding for the question
     query_vec = embedder.embed([question])[0]
-    top_docs = vector_store.search(query_vec, k=k)
 
-    # Deduplicate context
-    seen = set()
-    unique_docs = []
-    for doc in top_docs:
-        if doc not in seen:
-            seen.add(doc)
-            unique_docs.append(doc)
+    # Get top-k results with scores
+    results = vector_store.search_with_scores(query_vec, k=k)  # returns list of (doc, score)
+
+    # Filter by similarity threshold
+    top_docs = [doc for doc, score in results if score >= min_similarity]
 
-    context_text = "\n\n".join(unique_docs)
+    if not top_docs:
+        # No relevant context found
+        return "I’m sorry, I don’t have relevant information in my knowledge base for this query."
 
-    # Build clear, strict prompt
+    # Join context for the prompt
+    context_text = "\n\n".join(top_docs)
+
+    # Build prompt for TinyLlama
     prompt = f"""
 You are a helpful IT support assistant.
 
@@ -73,4 +74,6 @@ User question: {question}
 Answer:
 <|assistant|>
 """
+
+    # Generate answer
     return llm.generate(prompt)