Chat in explore is a conversational application used to explore the boundaries of Cyberwisdom TalentBot LLMops's capabilities.
When we talk to large natural language models, we often encounter situations where the answers are outdated or invalid. This is due to the old training data of the large model and the lack of networking capabilities. Based on the large model, Chat uses agents to capabilities and some tools endow the large model with the ability of online real-time query.
Chat supports the use of plugins and datasets.
LLM(Large language model)cannot be networked and invoke external tools. But this cannot meet the actual usage scenarios, such as:
When we want to know the weather today, we need to be connected to the Internet.
When we want to summarize the content of a web page, we need to use an external tool: read the content of the web page.
The above problem can be solved by using the agent mode: when the LLM cannot answer the user's question, it will try to use the existing plugins to answer the question.
In Cyberwisdom TalentBot LLMops, we use different proxy strategies for different models. The proxy strategy used by OpenAI's model is GPT function call. Another model used is ReACT. The current test experience is that the effect of GPT function call is better. To know more, you can read the link below:
Currently we support the following plugins:
Google Search. The plugin searches Google for answers.
Web Reader. The plugin reads the content of linked web pages.
Wikipedia. The plugin searches Wikipedia for answers.
We can choose the plugins needed for this conversation before the conversation starts.
If you use the Google search plugin, you need to configure the SerpAPI key.
Chat supports datasets. After selecting the datasets, the questions asked by the user are related to the content of the data set, and the model will find the answer from the data set.
We can select the datasets needed for this conversation before the conversation starts.
The process of thinking
The thinking process refers to the process of the model using plugins and datasets. We can see the thought process in each answer.