sunny-green•15mo ago
Is it possible to customize VAPI's RAG implementation?
Is it possible to customize VAPI's RAG implementation without using make.com, to avoid high latency? Specifically, can we integrate our own vector database for semantic and keyword search, use Unstructured for document processing, and implement reranking with Cohere, while maintaining low latency?
4 Replies
@AIMinato
Absolutely, you can customize VAPI's RAG (Retrieval-Augmented Generation) implementation to suit your specific needs and optimize for lower latency. Integrating your own vector database for semantic and keyword search is a great way to maintain control over the search performance and relevance. For document processing, using Unstructured can help you extract and process information efficiently. Lastly, implementing reranking with Cohere can enhance the quality of the results by ensuring the most relevant information is presented first.
To achieve this, you would typically need to:
1. Set up your own vector database that is optimized for fast retrieval. This could be something like Elasticsearch, FAISS, or any other efficient vector search solution.
2. Integrate Unstructured into your data processing pipeline to handle the extraction and processing of information from documents.
3. Use Cohere's reranking capabilities to reorder the results based on relevance, which can be done by scoring the results and sorting them accordingly.
By handling these components on your own infrastructure or choosing cloud services with low-latency characteristics, you can minimize the latency that might be introduced by external platforms like make.com.
For detailed guidance on how to implement these customizations with VAPI, you might want to check out the Advanced Concepts section in the documentation or reach out to the support team for more specific instructions. If you need further assistance or have any more questions, please let me know!
You can continue this conversation by mentioning me in the message.
ratty-blush•15mo ago
You can use tool call to integrate your vector db.
sunny-greenOP•15mo ago
I will try thx
ratty-blush•15mo ago
Closing then as you will try it later.