utils/llama_index_utils.py

import os
import json
from transformers import pipeline
from llama_index import SimpleDirectoryReader, GPTVectorStoreIndex, LLMPredictor, OpenAI

_index = None

def query_symptoms_tool(prompt_json: str):
    # parse “prompt_json” into Python dict and call your existing query_symptoms()
    data = json.loads(prompt_json)
    return query_symptoms(data["raw_input"])

def get_llm_predictor():
    """
    Return an LLMPredictor configured for local GPU (transformers) if USE_LOCAL_GPU=1,
    otherwise uses OpenAI.
    """
    if os.getenv("USE_LOCAL_GPU") == "1":
        # Local GPU inference using GPT-2 as an example
        local_pipe = pipeline("text-generation", model="gpt2", device=0)
        return LLMPredictor(llm=local_pipe)
    # Default to OpenAI provider
    return LLMPredictor(llm=OpenAI(temperature=0))


def build_index(data_path="data/icd10cm_tabular_2025"):  # noqa: C901
    """
    Build (or retrieve cached) GPTVectorStoreIndex from ICD documents.
    """
    global _index
    if _index is None:
        # Load documents from the ICD data directory
        docs = SimpleDirectoryReader(data_path).load_data()
        # Initialize the index with chosen LLM predictor
        predictor = get_llm_predictor()
        _index = GPTVectorStoreIndex.from_documents(docs, llm_predictor=predictor)
    return _index


def query_symptoms(prompt: str, top_k: int = 5):
    """
    Query the index for the given symptom prompt and return the result.
    """
    idx = build_index()
    # Create a query engine with the same predictor
    predictor = get_llm_predictor()
    query_engine = idx.as_query_engine(similarity_top_k=top_k, llm_predictor=predictor)
    return query_engine.query(prompt)