Was looking into applying LLMs onto entity extraction in earnings calls. (Around 400mb of text data). 1 Million tokes is around 1-3$.<p>Open for ideas and experiments.
Just be careful with the "temperature" when dealing with numbers and financial data as well. You should check samples of the retrieved data by hand.<p>Setting the temperature lower than 0.9 means reducing the "creativity" and making it less prone to hallucinate.
That's not how you do it. You use a Vectordatabase and a Retriever. This way, not all tokens of a document are used with the prompt, just the relevant parts will end up in the conversation / prompts. This way you save a lot of money and you are NOT limited to use GoogleAI - you can use whatever AI you want.
Use Flowise and a VectorDB like upstash or pinecone.<p>Flowise Tutorial
<a href="https://www.youtube.com/watch?v=V7uBy3VQJAc" rel="nofollow">https://www.youtube.com/watch?v=V7uBy3VQJAc</a>
Can’t you pre-filter it with a vector lookup or something cheap first (gpt3.5?). Or even simply filter first by other things you are considering for the trading decision (price to earnings for example). That might save you a lot of cost.