I want to build a conversational chatbot which have knowledge base as well as universal knowledge of GPT-3

I want to build a conversational chatbot which have knowledge base as well as universal knowledge of GPT-3. So, I want GPT-3 to answer Questions/Queries/Conversation based on knowledge base (sort of QA) and When there is nothing relevant in Knowledge base it will answer from last 10-12 conversation History. My knowledge base have few pdf docs which is long. So, It is not possible to send it all the time to GPT-3. Is there anyway I can use my old knowledge base for all the conversations without sending in all the prompts to save cost?


I don’t know of a way to avoid paying for all tokens in the knowledgebase if you want GPT-3 to search within it. The max_rerank can help lower costs, but as was mentioned in one of the other threads, it causes the first search to be done via keyword-match, nullifying some of the benefits of a full semantic search.

If you want knowledge-base, followed by general-knowledge, you can daisy chain. If the general-knowledge query returns null, re-ask the query to the general-knowledge endpoint

It is not lot like a Question Answer but it have context of few knowledge based summaries, definitions and details etc… I just want to keep that context common throughout the conversation as well as reduce the cost.

Like @m-a.schenk mentioned about, you may want to read up on the Fine Tuning beta. Since its new, we need to work out the math to see if its cheaper than the traditional pricing model


I am unable to find how to upload pdfs in finetuning and also found few blogs saying finetuning is not for everyone. I don’t have any pairs of conversations i have is some definations and summaries in document. How can I use that for finetuning can you provide me some reference or something.

1 Like

I am about to publish my book on this topic. I will update this thread once it’s out. I’m aiming for this coming weekend.

EDIT: David K Shapiro - Natural Language Cognitive Architecture


It’s not very straightforward to directly upload pdfs and run inference/finetune model from their contents. The reason being that pdfs can contain all kinds of data (images, tables, latex etc) and it all has to be normalized and cleaned.

Take a look at the haystack project. It is a toolkit for building pipelines for q&a, search, qg etc tasks and it’s fairly easy to get started. They have a node specifically for using GPT3 for the answer generation.