The C3 AI Code Assistant, trained on the technical ins and outs of the C3 AI Platform and applications, will accelerate customer support in our online forum for data scientists, engineers, and developers

By Ishaan Gupta, Software Engineer, Jeet Vachhani, Software Enginner, Harsha Vaddi, Data Scientist, Semi Hasaj, Senior Data Scientist, and Amir Delgoshaie, Director, Data Science, C3 AI

Advancements in generative AI have made it easier than ever to provide exceptional technical customer support that is fast and available 24/7. Using these capabilities, we are excited to announce the general availability of C3 Generative AI for our online developer community: the C3 AI Code Assistant.

The C3 AI Community is an online engagement forum for C3 AI users and developers. Here, you can post comments and questions about all C3 AI products or get support developing new AI applications on the C3 AI Platform. These questions can cover a wide range of topics — from how to build a faster machine learning pipeline, how to configure a dashboard in your application, or even troubleshooting user access and privileges. Previously, only C3 AI experts responded to these questions. Now, Code Assistant has joined the Community to help resolve questions faster.

Code Assistant delivers answers to user questions in seconds

The C3 AI Code Assistant is an agent that uses a specialized multimodal LLM trained on the C3 AI Platform and application codebases. It is tuned to respond as a helpful expert by absorbing thousands of feedback instances from our internal and external developers. The Code Assistant has access to powerful tools like document search, code generation, memory, syntax checker, grounding, and hallucination guardrails. Code Assistant can extract the necessary information from your questions and screenshots, plan how best to help you resolve your problem using various tools, generate and validate code examples specific to your questions, and share valid links to relevant documentation.

As soon as users post a new question on the C3 AI Community, they receive a helpful response from the Code Assistant within seconds. Our analysis shows that in most cases, Code Assistant provides a response that resolves the question or points the user in the right direction by sharing relevant documentation. This has significantly reduced the time needed to resolve questions in our developer community, in many cases from hours to seconds.

Code Assistant uses images and screenshots from user questions to identify errors and provide solutions

Develop Faster by Getting an Instant Response from a C3 AI Expert

Here are some of the key features of the C3 AI Code Assistant in the Community:

Grounded in Documentation

All responses from the C3 AI Code Assistant are grounded in documentation to provide a factually correct response and ensure easy access to all available resources related to their question. Moreover, the retrieved documentation is from the exact product version mentioned in the post, so that we provide precise code examples and documentation from that exact version.

Minimized Hallucinations

Our assistant uses a hallucination guardrail that leverages an LLM as a judge to ensure the response does not include irrelevant or incorrect information. Our experiments show that for C3 AI’s domain-specific language, our specialized LLM judge outperforms best-in-class, general-purpose models. For complex questions, Code Assistant will ask for additional information and wait for guidance from an expert rather than providing an incorrect response.

Conversation-Aware Responses

Many support questions are complex and require input from C3 AI experts. In these cases, the C3 AI Code Assistant will follow the flow of the conversation and respond if prompted by one of the participants.

Code Assistant understands the context of a thread and can provide follow-up responses when prompted by users

Special Attention to Code

Importantly, the C3 AI Code Assistant’s tools include code checks to ensure that responses containing code snippets are safe and syntactically correct. We found that a few seconds of extra wait time for the user was a good tradeoff for higher quality responses that we can get from agents that validate and correct the LLM-generated code before presenting it to the end user.

Powered by C3 Generative AI

We have built this powerful system on top of C3 Generative AI. If you need to build an assistant to make your developer support significantly more efficient, consider using the powerful building blocks available in C3 Generative AI. Learn more about how to get started with our C3 Generative AI Accelerator Program.

If you’re a customer, join us in the C3 AI Community to learn more and enhance your AI projects.