llm_services/tools.py

import os
from openai import OpenAI
import logging
from typing import List, Dict, Any
import json
import time
from pydantic import BaseModel
import time
from qdrant_client import QdrantClient, models
from qdrant_client import QdrantClient
from qdrant_client.models import Filter, FieldCondition, MatchValue, Range, MatchAny


qdrant_client = QdrantClient(
    url="https://00e40cf4-6976-43c1-aa08-be895735804b.europe-west3-0.gcp.cloud.qdrant.io:6333",
    api_key="eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJhY2Nlc3MiOiJtIn0.lfvuJEBzsmB7vez0nhv7HBbUlW77eUAT8raSazqYXHA",
)

qdrant_client.set_model("sentence-transformers/all-MiniLM-L6-v2")
qdrant_client.set_sparse_model("prithivida/Splade_PP_en_v1")

def recommend_talent_tool(search_query: str, verticals=None, min_age=None, max_age=None, gender=None,
                          min_followers=None, max_followers=None,
                          min_overall_engagement=None, max_overall_engagement=None):
    conditions = []

    if verticals:
        if isinstance(verticals, str):
            verticals = [verticals]
        conditions.append(
            FieldCondition(
                key="verticals",
                match=MatchAny(any=verticals)
            )
        )

    if min_age is not None or max_age is not None:
        range_params = {}
        if min_age is not None:
            range_params["gte"] = int(min_age)  # Ensure it's an integer
        if max_age is not None:
            range_params["lte"] = int(max_age)
        conditions.append(
            FieldCondition(
                key="age",
                range=Range(**range_params)
            )
        )

    if gender:
        conditions.append(
            FieldCondition(
                key="gender",
                match=MatchValue(value=gender)
            )
        )

    if min_followers is not None or max_followers is not None:
        range_params = {}
        if min_followers is not None:
            range_params["gte"] = int(min_followers)  # Convert to int
        if max_followers is not None:
            range_params["lte"] = int(max_followers)
        conditions.append(
            FieldCondition(
                key="follower_count",
                range=Range(**range_params)
            )
        )

    if min_overall_engagement is not None or max_overall_engagement is not None:
        range_params = {}
        if min_overall_engagement is not None:
            range_params["gte"] = float(min_overall_engagement)  # Convert to float
        if max_overall_engagement is not None:
            range_params["lte"] = float(max_overall_engagement)
        conditions.append(
            FieldCondition(
                key="overall_engagement",
                range=Range(**range_params)
            )
        )

    query_filter = Filter(must=conditions) if conditions else None

    search_result = qdrant_client.query(
        collection_name="social_media_profiles",
        query_text=search_query,
        query_filter=query_filter,
        limit=10
    )

    # Extract metadata and scores
    results = []
    max_followers = max((hit.metadata.get("follower_count", 1) for hit in search_result), default=1)
    max_engagement = max((hit.metadata.get("overall_engagement", 1) for hit in search_result), default=1)

    W1, W2, W3 = 0.5, 2.5, 0.25

    for hit in search_result:
        metadata = hit.metadata
        vector_similarity = hit.score  # Qdrant should return a similarity score
        
        follower_count = metadata.get("follower_count", 0)
        overall_engagement = metadata.get("overall_engagement", 0)

        # Normalize follower count and engagement score
        normalized_followers = follower_count / max_followers if max_followers > 0 else 0
        normalized_engagement = overall_engagement / max_engagement if max_engagement > 0 else 0

        # Calculate the matching score
        matching_score = (
            W1 * vector_similarity +
            W2 * normalized_followers +
            W3 * normalized_engagement
        )

        results.append({"metadata": metadata, "matching_score": matching_score})

    # Sort by matching score in descending order
    sorted_results = sorted(results, key=lambda x: x["matching_score"], reverse=True)

    return [item["metadata"] for item in sorted_results]