How Scriv Works#

Understanding how Scriv works is useful for building a mental model for the things it is good—and not good—at. It’s also helpful for understanding how you can get good answers from your bots.

Scriv answers questions using a process called retrieval augmented generation, or RAG for short.

Retrieval augmented generation is the process of supplementing a user’s input to a large language model (LLM) like ChatGPT with additional information that you have retrieved from somewhere else. The LLM can then use that information to augment the response that it generates.

This diagram shows how it works in practice:

RAG Overview

It starts with an incoming question. For example “How do I do ?” The first thing that happens is the retrieval step. This is the process that takes that question and searches for the most relevant content from a knowledge base that might answer it. You can think of this part like a search engine for your uploaded content.

After retrieval, the relevant information is sent, along with the question, to the LLM (ChatGPT or GPT4). The LLM reads the information and then uses it to construct and answer.

What’s important to note is that only the relevant information from your knowledge base gets included in the answer. The entire context is never fully available.

What is Scriv good at?#

Scriv is very useful when you have a large knowledge base—e.g. one or more documentation sites, PDF documents, internal wikis, or Slack communities—and want to ask specific questions.

If you think the answer is somewhere in that information but aren’t sure where, Scriv can find it and answer precisely, along with citing its sources in case you wnat to double-check its work.

This makes it good for use cases like:

  • Providing product support (either for support teams or direct to customers)

  • Internal knowledge resources, for example, bots for employees based on company wikis

  • Research projects, e.g. looking for answers in a collection of scientific papers

  • Developer tools, based on documentation sites.

What is Scriv bad at?#

Because Scriv is always only looking at a subset of the knowledge base, it’s not good for tasks that require accessing the whole thing.

So it is bad for use cases like:

  • Generating summaries.

  • General questions—e.g. “what is the content about”?

  • Analyzing and running calculations on a spreadsheet.

One helpful rule of thumb is that questions/answers that are in the content are good, but questions/answers about the content are less likely to work.

Making the most of Scriv#

Here are some guidelines for getting the most from Scriv:

Be as specific as possible#

Since the question is used as a search text, it’s useful to make it as specific as possible.

Bad: Summarize the material

Good: Summarize how I can get the most out of using Scriv

Don’t put instructions in the question#

Again, since the question is used as search text, the instructions can throw it off.

Bad: What knowledge sources can I use? Please answer as a list, using Markdown.

Good: What knowledge sources can I use?

In above example, if you want the bot to use lists with Markdown, you can add those as custom instructions to your bot.