I have a two examples of chabot implementation where in first one i use GPT-3-turbo and in the second GPT-4-turbo. I am getting better results from GPT-3-turbo than from GPT-4-turbo. Context: I use Pinecone VectorDB storage and query the data using query engine. I did switch to a new embedding model like so :
"llm = OpenAI(model="gpt-4-turbo", temperature=0)
Settings.llm = llm
embed_model = OpenAIEmbedding(model="text-embedding-3-small")
Settings.embed_model = embed_model
#logging.info(" LLM MODEL OPENAI" + llm.model)
logging.info("Initialized OpenAI")
query_engine = loaded_index.as_query_engine(streaming=False) #, text_qa_template=text_qa_template, llm=llm)
logging.info("Initialized QueryEngine")"
The second more simpler implementation of my bot :
# Initialize your index
pinecone_index = pc.Index(index_name)
vector_store = PineconeVectorStore(pinecone_index=pinecone_index)
loaded_index = VectorStoreIndex.from_vector_store(vector_store=vector_store)
logging.info("Index loaded from Pinecone")
query_engine = loaded_index.as_query_engine()Is the embedding model problematic ? I used llamacloud parse and then indexed the data... I am using the newer embeding model "text-embedding-3-small" in the GPT-4-turbo version. Could this be making my GPT-4-turbo results worse ? I do have langfuse connected to better analyze the retrieve process.
