service/rag_service.py

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

# -------------------------------
# Setup
# -------------------------------

CACHE = Path("embeddings.pkl")

# Embedder and LLM
embedder = EmbeddingService()
llm = LLMService()

# Load documents
loader = CSVDataLoader("final_data_set(in).csv")
documents = loader.load_qa_pairs()

# Load or compute embeddings
if CACHE.exists():
    with CACHE.open("rb") as f:
        embeddings = pickle.load(f)
else:
    embeddings = embedder.embed(documents)
    with CACHE.open("wb") as f:
        pickle.dump(embeddings, f)

vector_store = VectorStoreService(embeddings, documents)

# -------------------------------
# Generate answer function
# -------------------------------

def generate_answer(question: str, k: int = 3, min_similarity: float = 0.65) -> str:
    """
    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]

    # 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]

    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."

    # 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.

- ONLY answer based on the context below.
- DO NOT hallucinate or invent new steps.
- If the answer is not explicitly present in the context, return the context itself.
- Keep answers concise and step-by-step if possible.

Context:
{context_text}

User question: {question}

Answer:
<|assistant|>
"""

    # Generate answer
    return llm.generate(prompt)