The solution/concept you are looking for is Retrieval-Augmented Generation (RAG) for LLMs. If you want to learn the concept, I like the llamaindex documentation
Conceptually, RAG is similar to going to a search engine and grabbing all the relevant paragraphs for a llm to summarize.
If you want a sample to play around with, I recommend playing around with this code snippet
This guide uses the transformers library, but I use Ollama with the OpenAI api.
If you opt for the build, the biggest challenge is taking your internal wikis and documents and getting them ready for a database and a database solution. If you want a all in one solution, I’d check out Trieve. They have some cool super fast demos.
They also have a version you can self host
Disclaimer I got to chat with the trieve devs they know their stuff and have some super fast demos built on their tech
I also found them through the FUTO channel
After you get the chatbot part complete, you’ll need to hook it into a speech to text and text to speech solution. I’m not as familiar with this space but look for open source TTS and STT models
If your company is on Microsoft and Azure you might want to consider AI search and AI speech or the AWS and GCP equivalents. It will cost you, but if your leaders are more comfortable working big tech, then it might be worth looking into
Google has probably the most generous free testing for developers