CRM Data enrichment

[sanmbitbiswas]

Many thanks to the creators for this project. It has been a life saver for me.

I'm involved with enriching data on missing fields in CRM as an ops person. I generally use the Perplexica API to find answers to simple questions like - find the website of a company, list the top three services of a company, find LinkedIn URL of company etc, a person's LinkedIn URL etc.

In most cases I get correct answers but there are times when the answers are incorrectly generated. Eg - Urls in the answer don't work, incorrect information about the company etc.

I'm using Ollama models, nomic embed for embedding, and llama3.2 instruct model.

Can anyone suggest how I can improve the quality of answers.

I have a 12 gb vram nvidia rtx 4070 GPU with 32 gb ram. Are those specs enough to generate good answers?

Should I use a bigger local model like Gemma 12 B?

Is there any parameter I can tweak?

Thank you in advance for any helpful suggestion

[George1422]

There are a couple of things to unpack here.

1. What quality of response are you expecting to get that llama3.2 can't give you if you are asking just simple find me, get me,
2. From a size perspective Gemma3:12b will fit into memory on the 4070 with room for the context window. So you can upgrade if you want. I would at least try it to see if it nets you any better answer. Just remember that bigger is slower.

FWIW Gemma3:4b fits into 3.5GB of vram. You didn't happen to mention the number of parameter for llama3.2 IMO llama3.2 is a totally capable model but it is several years old in its training. Are you using llama for its function calling capability, if not take a peek at gemma3. Simple question and answer can be handled by the smaller llms, if you want creativity and robustness in response use the larger llm, knowing your response times will increase because of the additional capabilities.

I also have to wonder if perplexica is the right tool for your task? Again this is just a personal opinion, I wonder if RAG with your own documents would be a different solution. If you use open web-ui and loaded all of your company information into a rag document(s) store then posted questions to the llm referring this document store. Would it give you a better quality of answer(??). Understand this method would not search the web to give you an answer but would only use the information you told it (in this case the llm would only need to understand your intent, search the document store, and then formulate a response to the question. If you are using an external api to access perplexica you could also do that with the open web-ui system too. Again I'm basing that opinion on what you have stated as your difficulties with perplexica, there may be other considerations you haven't mentioned.