SonicSpec helps you save time reviewing specification line items ("spec phrases") by giving you the ability to store a compliance response and quote consideration with an individual spec phrase so that you can access these responses the next time the same spec phrase is seen again in a spec uploaded to SonicSpec.

But what can be done about spec phrases that your organization has not previously reviewed before?

By default, SonicSpec will provide suggested responses by displaying your saved compliance response and quote consideration to the most similar spec phrase in your library. Often, these will be the correct responses and you will save time by simply accepting the suggested responses rather than typing them in manually.

However, there is one severe limitation to these base-level suggestions: they are not "smart".

For instance, take a look at this suggestion:

You can see that the SonicSpec suggestion engine found a very similar spec phrase that had been reviewed before. Structurally, these two spec phrases are extremely similar. Unfortunately, they differ in the most critical aspect: the altitude at which the products must be rated. Missing this one dimension could lead to a critical error. Therefore, without any reasoning capability, SonicSpec's base-level suggestions cannot be trusted as a second set of eyes, and you should only use these suggestions to help you input responses that you've already evaluated faster by saving keystrokes on your keyboard.

SonicSpec offers integrations with the leading large language models, OpenAI and Anthropic, to insert reasoning capabilities into your suggestions so SonicSpec can be your second set of eyes on a spec review whether you've previously evaluated the spec phrases or not.

Using the AI-Powered Suggestions button, we've successfully added the reasoning necessary to fix our previously incorrect suggestion:

AI-Powered Suggestions leverage your company's specification library, your base product prompt (i.e. instructions), and your company's knowledge base to send all the relevant information to a large language model that is needed to generate a smart, well-reasoned suggestion.

As you review specifications in SonicSpec, you are building up your company's specification library. Every time you add a compliance response or quote consideration to a spec phrase, you are saving this spec phrase/response pair to your spec library. The more you do this, the more powerful your spec library becomes.

Previously-evaluated spec phrases can be very helpful for providing the context to a language model needed to generate an accurate response.

For example, perhaps you have reviewed a fuel tank specification that required a 96 hour capacity, and you responded "comply." Using this example, a language model can confidently assess a 72 hour requirement by taking the fact that your company can provide a 96 hour fuel tank and respond "comply" to the 72 hour requirement as 96 hours is more restrictive than 72 hours.

Your prompt is the next step towards configuring your AI-Powered Suggestions. Your prompt provides instructions to the large language model about how to generate the text you want: a well-reasoned compliance response or quote consideration. Large language models are general-use tools that can help you with an infinite number of tasks. ChatGPT is the most well-known language model interface. In simple terms, you can think of your prompt as the equivalent to writing your chat message in ChatGPT if you were trying to get ChatGPT's help in reviewing a spec phrase. If you wanted ChatGPT to help you evaluate a spec phrase, you need to give it all the relevant information first. You need to clearly instruct the language model about what its job is, how it must accomplish the job you expect from it, and what its expected output is.

A few suggestions for structuring your prompt:

SonicSpec offers a few shortcuts that you can insert into your prompt to automatically retrieve relevant information for a particular spec phrase. You can think of this as spec-phrase-specific information. For example, you may want installation information added to your prompt only when evaluating an installation-related spec phrase. You can utilize the following context calls (with the curly brackets) in your prompt:

The {spec_phrase} context call pulls in the spec phrase being evaluated. Therefore, you must utilize this context call if you want the language model to evaluate your spec phrase. This is often done towards the end of the prompt after providing all of your instructions and context.

The {context} context call will automatically pull in three similar spec phrases from your library to the one being evaluated, as well as the saved responses for each one. This often provides great context for the language model to draw from as it attempts to interpret the spec phrase at hand, and also provides a great stylistic guide for how you want your compliance response or quote consideration to look.

The {knowledge_base} context call will retrieve the most relevant "chunk" from your knowledge base (more on this below). If your specification discusses fuel tanks and you have a chunk dedicated to fuel tank information, SonicSpec will automatically identify this chunk as the most relevant and pull it into your prompt for the given spec phrase. This allows you to directly insert relevant product data, instructions, etc. while managing the amount of information you add to each prompt. This helps you get more accurate responses as a higher percent of your prompt is relevant to the individual task, while also helping you control costs as AI-Powered Suggestions are billed based on the number of words (tokens) sent to and from the language model.

Your knowledge base is a repository of product data that can be automatically inserted into a given prompt to provide the language model relevant information as you ask it to evaluate a never-before-seen spec phrase.

You can access and add to your knowledge base from your "Library" tab:

You can add product data, large language model instructions, or anything you think would help provide a better spec phrase interpretation by the language model.

When you use the {knowledge_base} context call (as described above), SonicSpec will automatically pull the most relevant knowledge base chunk into a prompt it identifies as being the most similar to the given spec phrase that is being evaluated. For example, if a spec phrase is discussing submittal requirements, it will look to find a knowledge base chunk that discusses submittal spec interpretations. Similarly, if the spec phrase is discussing auxiliary contact requirements, it will look to find a chunk that discusses auxiliary contact product data and instructions.

There are no limitations on how you choose to break down your knowledge base into chunks. However, we recommend that you try to keep each chunk somewhat relevant to a specific topic that is often referenced in specifications. If you have some information about your fuel tank capabilities in one chunk and some other information about your fuel tank capabilities in another chunk, you may not be able to predict whether the right fuel tank chunk is inserted into the prompt that is evaluating a fuel tank spec phrase.

Getting Help

We know that learning how to craft prompts and create knowledge bases can be a daunting task. We are here to help!

Reach out to support@sonicspec.com to get customized help on building AI-Powered Suggestions that work.