# query_topk_prompt.py
# Run a querry on a vector store
# 
# This version from query_rewrite_hyde.py, but removing hyde and using a custom prompt
# This verison implements a prompt and uses the build_exp.py vector store with BAAI/bge-large-en-v1.5
# Based on query_exp.py->query_topk.py->query_rewrite_hyde.py
# The results are as good as with HyDE.
#
# E.M.F. August 2025

from llama_index.core import (
    StorageContext,
    load_index_from_storage,
    ServiceContext,
    Settings,
)
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
from llama_index.llms.ollama import Ollama
from llama_index.core.prompts import PromptTemplate
import os

#
# Globals
#
os.environ["TOKENIZERS_PARALLELISM"] = "false"

# Embedding model used in vector store (this should match the one in build_exp.py or equivalent)
# embed_model = HuggingFaceEmbedding(model_name="all-mpnet-base-v2")
embed_model = HuggingFaceEmbedding(cache_folder="./models",model_name="BAAI/bge-large-en-v1.5")

# LLM model to use in query transform and generation
# command-r7b generates about as quickly as llama3.1:8B, but provides results that stick better
# to the provided context
llm="command-r7b"
# Other models tried:
#llm="llama3.1:8B"
#llm="deepseek-r1:8B"
#llm="gemma3:1b"

#
# Custom prompt for the query engine
#
PROMPT = PromptTemplate(
"""You are an expert research assistant. You are given top-ranked writing excerpts (CONTEXT) and a user's QUERY.

Instructions:
- Base your response *only* on the CONTEXT.
- The snippets are ordered from most to least relevant—prioritize insights from earlier (higher-ranked) snippets.
- Aim to reference *as many distinct* relevant files as possible (up to 10).
- Do not invent or generalize; refer to specific passages or facts only.
- If a passage only loosely matches, deprioritize it.

Format your answer in two parts:

1. **Summary Theme**  
   Summarize the dominant theme from the relevant context in a few sentences.

2. **Matching Files**  
   Make a list of 10 matching files. The format for each should be:  
   <filename> - 
   <rationale tied to content. Include date or section hints if available.>

CONTEXT:
{context_str}

QUERY:
{query_str}

Now provide the theme and list of matching files."""
)

#
# Main program routine
#

def main():
    # Use a local model to generate -- in this case using Ollama
    Settings.llm = Ollama(
        model=llm,    # First model tested
        request_timeout=360.0,
        context_window=8000
        )
   
    # Load embedding model (same as used for vector store)
    Settings.embed_model = embed_model

    # Load persisted vector store + metadata
    storage_context = StorageContext.from_defaults(persist_dir="./storage_exp")
    index = load_index_from_storage(storage_context)

    # Build regular query engine with custom prompt
    query_engine = index.as_query_engine(
        similarity_top_k=15,            # pull wide 
        #response_mode="compact"        # concise synthesis
        text_qa_template=PROMPT,        # custom prompt
        # node_postprocessors=[
        #     SimilarityPostprocessor(similarity_cutoff=0.75)  # keep strong hits; makes result count flexible
        # ],
    )  

    # Query
    while True:
        q = input("\nEnter a search topic or question (or 'exit'): ").strip()
        if q.lower() in ("exit", "quit"):
            break
        print()

        # Generate the response by querying the engine
        # This performes the similarity search and then applies the prompt
        response = query_engine.query(q)

        # Return the query response and source documents
        print(response.response) 


        print("\nSource documents:")
        for node in response.source_nodes:
            meta = getattr(node, "metadata", None) or node.node.metadata
            print(f"{meta.get('file_name')} {meta.get('file_path')} {getattr(node, 'score', None)}")


if __name__ == "__main__":
    main()