Skip to content

Creating a Research Assistant

Basic Configuration

Creating a new Research Assistant is incredibly simple.

  1. Simply click, "Create a Research Assistant" in the far left box.
  2. Give your Research Assistant a name and description.
  3. Choose the Library connections you want the Research Assistant to draw from.

Advanced Configuration

Sharing

Share With Organization

Share the Research Assistant with other members of your organization.

Your Research Assistants can be publicly shared with users outside of the platform. By enabling this setting, you will be provided with a link that you can then share with others.

Warning

Be careful about connecting libraries of sensitive information as this Research Assistant will be availiable to the general public.

Prompts

Adjusting the prompts, known as prompt engineering, is a powerful way to refine the answers that you get back from the Research Assistant.

Chat History Prompt

The chat history prompt is used by the research assistant to reformulate your questions based on your chat history. Importantly, it allows the chatbot to remember the context of the previous conversation. So, if you ask a question about Kamala Harris, and later in the chat refer to her as Kamala or Harris, it will know who you are referring to.

The default chat history prompt is the following:

Below is a conversation between a research assistant and a researcher. Your 
task is to use this conversation to rewrite  the final question posed by the
researcher. The rewritten  question should be clear and specific enough for
another  research assistant to understand and answer without 
needing to ask for additional context.

If you change this setting, the chat history prompt will be replaced with your custom prompt.

Tip

Most Research Assistant tuning can be accomplished by updating the Additional Instructions parameter.

Additional Instructions

The additional instructions parameter is a way to add additional instructions to your research assistant. This is a great opportunity to get creative with customizing your research assistant.

Some of our favorite examples include:

  • Return all answers in Spanish: Most LLMs can provide responses in a number of languages, however we recommend throrough testing by native speakers prior to deployment.

  • All of your answers should be short and concise: Depending on the output that you are seeking, you can specify whether answers should be concise or verbose.

  • Return all answers at a 7th grade reading level: Specify the reading level to ensure that it matches your audience.

  • Your tone should be excited and enthusiastic: Specify the tone of the responses, including neutral and academic.

  • Include citations with all responses: Specify the need for citations with all answers.

LLMs

By default, our platform uses OpenAI models, though we have the ability to leverage other LLMs if you so desire. These settings allow you to specify the specific LLMs used for your research assistants as well as mix and match for different aspects of the Research Assistant.

If you wish to utilize a non-OpenAI model, please contact [email protected].

Retrieval LLM

The LLM used to develop the question based on the chat history. Used with the chat history prompt.

Generation LLM

LLM used to formulate the answer based on the retrieved documents. Used with the chat prompt.

Retrieval

Max Documents Retrieved (K)

When you ask a question, the Research Assistant will search through your library for relevant chunks (subsets of your files) of documents to use to generate an answer. The default number of chunks that it retrieves is 3.

By increasing this parameter, it will retrieve more chunks. However, some of those documents may be less relevant, yield worse responses from the research assistant. It will also affect the speed with which it returns responses.

There are some instances in which you want to increase the maximum documents retrieved, such as when you are seeking answers that may be contained across a large number of documents. If you are asking questions like, "Tell me all of the times that Kamala Harris spoke about taxing the rich" then increasing this value may lead to better results.

Troubleshooting

You will be able to create a Research Assistant on any connection, even those that failed in their initial sync. Please make sure your connection has succeeded before using the Research Assistant in earnest. Once again, asking "What can I learn from the documents provided?" is a great way to ensure all of your intended documents are synced.

Tips & Tricks

Create Standalone Research Assistants for Clarity

Candidates

The language model can sometimes confuse individuals. If you are creating Research Assistants for the purpose of asking questions about a candidate and their opponent, it is best to create one Research Assistant for each candidate. So for example, you would create a Kamala Harris Research Assistant and a Donald Trump Research Assistant.

States

Because many Drive documents may make reference to "the state legislature" or "the Governor", the Research Assistant may confuse information across states. If you are using the tool for state-level work, it is more effective to create one Research Assistant per state.