llm_handler.py

# llm_handler.py
import google.generativeai as genai
from config import GOOGLE_API_KEY, GENERATIVE_MODEL, EMBEDDING_MODEL
import streamlit as st # For displaying errors or warnings if needed

# Configure Gemini API
if GOOGLE_API_KEY:
    try:
        genai.configure(api_key=GOOGLE_API_KEY)
    except Exception as e:
        st.error(f"Failed to configure Gemini API: {e}") # Show error in Streamlit if app is running
        print(f"Failed to configure Gemini API: {e}") # Print to console for server logs
else:
    # This will be handled by Streamlit UI in app.py if key is missing
    print("Warning: GOOGLE_API_KEY is not set. LLM features will not work.")


def get_gemini_response(prompt_text, system_instruction=None):
    """获取Gemini模型的响应"""
    if not GOOGLE_API_KEY:
        st.error("Gemini API Key未配置，无法获取模型响应。请在Hugging Face Space Secrets中设置 GOOGLE_API_KEY。")
        return None
    try:
        model = genai.GenerativeModel(
            GENERATIVE_MODEL,
            system_instruction=system_instruction if system_instruction else None
        )
        response = model.generate_content(prompt_text)
        return response.text
    except Exception as e:
        error_message = f"与Gemini通信时出错: {e}"
        if hasattr(e, 'message') and "API key not valid" in e.message:
             error_message = "Gemini API Key无效或权限不足。请检查Hugging Face Space Secrets中的GOOGLE_API_KEY。"
        st.error(error_message)
        print(error_message) # For server logs
        return None

# Using genai.embed_content directly is often simpler for ChromaDB
# but if you need a callable for ChromaDB's embedding_functions parameter:
class GeminiEmbeddingFunctionForChroma(genai.embedding.EmbeddingFunction):
    def __call__(self, input: genai.embedding.EmbedContentRequest) -> genai.embedding.EmbedContentResponse:
        # Ensure 'input' is a list of strings (documents)
        if not isinstance(input, list) or not all(isinstance(doc, str) for doc in input):
            # ChromaDB typically passes a list of documents (strings)
            # genai.embed_content expects a 'content' field which can be a string or list of strings
            # The structure of 'input' from ChromaDB needs to be correctly mapped.
            # ChromaDB's `embedding_function` interface expects a function that takes a list of texts
            # and returns a list of embeddings.

            # Let's assume 'input' is a list of document strings.
            docs_to_embed = input
        else: # Fallback if input structure is different, adapt as needed
            docs_to_embed = [str(item) for item in input]


        if not docs_to_embed:
            return {"embedding": []} # Return empty embedding list for empty input

        try:
            # Embed a batch of documents.
            # `task_type` is important for retrieval.
            result = genai.embed_content(
                model=EMBEDDING_MODEL,
                content=docs_to_embed,
                task_type="RETRIEVAL_DOCUMENT"
            )
            return result['embedding'] # ChromaDB expects a list of embeddings
        except Exception as e:
            error_message = f"获取文本嵌入时出错: {e}"
            st.error(error_message)
            print(error_message)
            # Return a list of Nones or empty lists of the correct length if an error occurs for some documents
            return [None] * len(docs_to_embed)

# --- Alternative simpler embedding function for ChromaDB ---
# This is often easier to integrate if ChromaDB's embedding_function
# parameter expects a function that takes a list of texts.
from chromadb import Documents, EmbeddingFunction, Embeddings

class GeminiChromaEF(EmbeddingFunction):
    def __init__(self, model_name: str = EMBEDDING_MODEL, task_type: str = "RETRIEVAL_DOCUMENT"):
        self._model_name = model_name
        self._task_type = task_type
        if not GOOGLE_API_KEY:
            print("Warning: GOOGLE_API_KEY not set. Embedding function might fail.")
            # Optionally raise an error or handle appropriately

    def __call__(self, input_texts: Documents) -> Embeddings:
        if not GOOGLE_API_KEY:
            st.error("Gemini API Key未配置，无法生成文本嵌入。")
            print("Gemini API Key not configured for embeddings.")
            return [([0.0] * 768) for _ in input_texts] # Return dummy embeddings or handle error

        if not input_texts:
            return []
        try:
            # Filter out any None or non-string inputs, though Documents type should be list of str
            valid_texts = [text for text in input_texts if isinstance(text, str)]
            if not valid_texts:
                # Handle case where all inputs were invalid
                return [([0.0] * 768) for _ in input_texts]


            result = genai.embed_content(
                model=self._model_name,
                content=valid_texts,
                task_type=self._task_type
            )
            # Ensure the result matches the number of valid_texts.
            # If there was an error, result['embedding'] might be shorter or None.
            # A robust handler would map results back to original input count, perhaps with None for errors.
            # For simplicity here, assuming success or a catastrophic failure handled by the try-except.
            
            # Map embeddings back to the original input_texts length, filling with None for invalid ones
            # This part is tricky because genai.embed_content might error out entirely or skip bad inputs.
            # Let's assume it returns embeddings for valid_texts only.
            embeddings_dict = {text: emb for text, emb in zip(valid_texts, result['embedding'])}
            
            final_embeddings = []
            for text in input_texts:
                if isinstance(text, str) and text in embeddings_dict:
                    final_embeddings.append(embeddings_dict[text])
                else:
                    # Provide a dummy embedding or None for invalid/missing inputs
                    # The dimension (e.g., 768) depends on your embedding model.
                    # For "models/embedding-001", it's 768.
                    final_embeddings.append([0.0] * 768) # Placeholder for invalid inputs
            return final_embeddings

        except Exception as e:
            error_message = f"获取文本嵌入时出错 (GeminiChromaEF): {e}"
            st.error(error_message)
            print(error_message)
            # Return dummy embeddings for all inputs in case of a general error
            return [[0.0] * 768 for _ in input_texts] # Placeholder dimension