src/qa.py

"""
qa.py — Retrieval + Generation Layer
-------------------------------------
Handles:
• Query embedding (SentenceTransformer / E5-compatible)
• Chunk retrieval (FAISS)
• Answer generation (Flan-T5)
Optimized for Hugging Face Spaces & Streamlit.
"""

import os
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
from vectorstore import search_faiss

print("✅ qa.py loaded from:", __file__)

# ==========================================================
# 1️⃣ Hugging Face Cache Setup (Safe for Spaces)
# ==========================================================
CACHE_DIR = "/tmp/hf_cache"
os.makedirs(CACHE_DIR, exist_ok=True)

os.environ.update({
    "HF_HOME": CACHE_DIR,
    "TRANSFORMERS_CACHE": CACHE_DIR,
    "HF_DATASETS_CACHE": CACHE_DIR,
    "HF_MODULES_CACHE": CACHE_DIR
})

# ==========================================================
# 2️⃣ Query Embedding Model
# ==========================================================
# Use E5-small-v2 for retrieval consistency with embeddings.py
try:
    _query_model = SentenceTransformer(
        "intfloat/e5-small-v2",
        cache_folder=CACHE_DIR
    )
    print("✅ Loaded query model: intfloat/e5-small-v2")
except Exception as e:
    print(f"⚠️ Query model load failed ({e}), falling back to MiniLM.")
    _query_model = SentenceTransformer(
        "sentence-transformers/all-MiniLM-L6-v2",
        cache_folder=CACHE_DIR
    )
    print("✅ Loaded fallback model: all-MiniLM-L6-v2")

# ==========================================================
# 3️⃣ LLM for Answer Generation (FLAN-T5)
# ==========================================================
MODEL_NAME = "google/flan-t5-base"   # switch to 'large' if RAM allows
print(f"✅ Loading LLM: {MODEL_NAME}")

_tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
_model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)

_answer_model = pipeline(
    "text2text-generation",
    model=_model,
    tokenizer=_tokenizer,
    device=-1  # CPU-safe for Spaces
)

# ==========================================================
# 4️⃣ Prompt Template (concise and factual)
# ==========================================================
PROMPT_TEMPLATE = """
You are an expert enterprise assistant.
Using ONLY the CONTEXT below, answer the QUESTION clearly and factually.
If the context doesn’t contain the answer, reply exactly:
"I don't know based on the provided document."

---
Context:
{context}
---
Question:
{query}
---
Answer:
"""

# ==========================================================
# 5️⃣ Chunk Retrieval Function
# ==========================================================
def retrieve_chunks(query: str, index, chunks: list, top_k: int = 3):
    """
    Encodes the user query and retrieves top-k relevant chunks via FAISS.
    Uses 'query:' prefix (E5 training style) for semantic alignment.
    """
    if not index or not chunks:
        return []

    try:
        # E5 expects 'query:' prefix for better retrieval accuracy
        query_emb = _query_model.encode(
            [f"query: {query.strip()}"],
            convert_to_numpy=True,
            normalize_embeddings=True
        )[0]

        results = search_faiss(query_emb, index, chunks, top_k)
        return results

    except Exception as e:
        print(f"⚠️ Retrieval error: {e}")
        return []


# ==========================================================
# 6️⃣ Answer Generation Function
# ==========================================================
def generate_answer(query: str, retrieved_chunks: list):
    """
    Generates an answer using FLAN-T5 and retrieved chunks as context.
    """
    if not retrieved_chunks:
        return "Sorry, I couldn’t find relevant information in the document."

    # Merge retrieved chunks for context
    context = "\n\n".join([f"[Chunk {i+1}]: {chunk}" for i, chunk in enumerate(retrieved_chunks)])

    # Build structured prompt
    prompt = PROMPT_TEMPLATE.format(context=context, query=query)

    try:
    result = _answer_model(
        prompt,
        max_new_tokens=350,        # allow longer, more complete answers
        do_sample=True,            # enable sampling for natural flow
        temperature=0.7,           # slightly higher = more expressive responses
        top_p=0.95,                # nucleus sampling for coherence
        repetition_penalty=1.2     # discourages repetitive phrasing
    )

    answer = result[0]["generated_text"].strip()

    # 🧩 If the model outputs something too short, expand gracefully
    if len(answer.split()) < 8:
        answer = (
            "The document mentions this briefly. Based on the context, here's what it suggests: "
            + answer
        )

    return answer

except Exception as e:
    print(f"⚠️ Generation failed: {e}")
    return "⚠️ Error: Could not generate an answer at the moment."



# ==========================================================
# 7️⃣ Optional Local Test
# ==========================================================
if __name__ == "__main__":
    dummy_chunks = [
        "SAP Ariba is a cloud-based procurement solution.",
        "It helps companies manage suppliers and sourcing processes efficiently.",
        "Integration with SAP ERP allows for seamless data synchronization."
    ]
    from vectorstore import build_faiss_index
    import numpy as np

    index = build_faiss_index([
        _query_model.encode([f"passage: {chunk}"], convert_to_numpy=True, normalize_embeddings=True)[0]
        for chunk in dummy_chunks
    ])

    query = "What is SAP Ariba used for?"
    retrieved = retrieve_chunks(query, index, dummy_chunks)
    print("🔍 Retrieved:", retrieved)
    print("💬 Answer:", generate_answer(query, retrieved))