Time to build: Creating your own company GPT

Why Build?

Artificial Intelligence is top of mind and, at times, already essential for businesses. Building your own company GPT—a chat assistant capable of accessing internal documents, searching personal files, collecting online information, and researching details from uploaded content—is now an easily achievable goal. While there is be a large cohort of startups lathered up to offer you this service, implementing it yourself is not overly difficult. One person with a series of helpful tools can get this type of tool running in 2-5 days.

So, let’s dive into the step-by-step process on how to make your company GPT while addressing common questions like “How do I comply with my company’s privacy policies?”

Understanding the practical steps of deploying THE classic AI use case

This is a practical guide to developing a working internal software product for your company. It is technical but has linked resources throughout to help AI beginners understand any relevant context.

Step 1: Choosing Your Foundation

Every good AI project starts with setting the right foundation. The key pieces of this foundation are the following 1. A researching agent 2. A hosting service 3. A database and 4. A language model. For this illustrative project, we will use the following: 1. GPT Researcher’s open source code as the structure for our research agents (website, code), 2. Vercel to host the app, Weaviate to store the data, and OpenAI’s GPT4o as the language model. These products are built specifically for AI development and enable fast development with a flexible foundation for future customization.

Step 2: Initial Setup and Troubleshooting

To get started with the development of this product, go to github and clone (read “download” for non-technical people) gpt-researcher as your starting code. This will serve as the logic foundation for your company GPT and has a series of well thought out prompts and workflows that collect meaningful information from your documents. This is a straightforward—"download and run"— process, however it requires some common adjustments and setup steps. One of these steps includes adding your environment variables to your code which allows you to access your specific instance of ChatGPT, Weaviate, and other products you use in the process. This is one of many small roadblocks you might face if you are setting up this type of system for the first time, but by using an AI companion like ChatGPT or even better an “in-code” editor like Cursor you will be able to copy and paste your error messages to the AI system and quickly solve any setup issues. These issues could look like needing to download a more recent version of code that is used to support gpt-researcher, changing your setup if the original code is written for a Mac and you run a PC, or getting permission from your system administrator to change files on your computer. After your troubleshooting is done, you should be able to follow the basic instructions from the GPT Researcher team and see your own “deep-research” web agent working on your computer in a web browser as shown in last minutes of this demo video.

Step 3: Integrating Complex Data Sources (Google Drive, Outlook)

After setting up your company GPT with web access through a service like Tavily, as described in the gpt-researcher instructions, the critical missing piece is your own personal data. There are a wide variety of processes that can be used to add this data but this is the general process:

1. Connect to Personal Data: Use data connectors like the Google Drive API, Microsoft Graph API, or the startups Ragie and Swirl to connect your data to a database.
2. Make Data Searchable: Convert the connected files into entries in a vector database. Services like Weaviate or Activeloop will allow you to convert and store your Google Drive data in a database that can be easily searched.
3. Store the Data: After making the vector database you now need to store the data. You can do this through Weaviate’s managed service or use a service like an AWS S3 bucket .

This process can seem overwhelming, so to simplify the initial setup, start with Google Drive by following the API documentation's setup instructions. Use a sample Drive folder and a Weaviate database for quick testing of the overall system.

Step 4: Connecting Across All Your Data Sources

You now have three search capabilities that can add information into your company GPT instance 1. Internet data 2. Company Data Across Apps 3. Direct File Uploads. The GPT researcher's original code will automatically integrate these sources by adding them to the context window of your language model. As you grow in complexity you may want to tweak how much and what type of information is pulled into these models.

At this stage of the project you have a research agent that can access these data sources:

1. Internet Data: Using a service like Tavily which is the default option in the gpt-researcher project, your company GPT now can search and integrate in web information.
2. Company Data Across Apps: With your “company data” that you have set up using Step 3, you will now be able to automatically pull in the most relevant data from your internal company files.
3. Direct File Uploads: In addition to the internal document search, you can also directly upload a document that you want the model to focus on. These documents will be directly referenced in the research output regardless of how relevant they are.

Step 5: Tying Everything Together For Your Company

To finalize the build of this company GPT you may want to adjust elements of the system to have the perfect set up for your organization. While these may take more time to understand they are often worth it to get the perfect report or answer out every time.

1. Selecting data combinations: After adding in the right document sources (e.g, your Google Drive), you can tweak what folders and how many documents you provide to your new company gpt. This will enable a more customized model that fits your organization or department needs best.
2. Depth of Analysis: Set the complexity and length of responses, from quick summaries to deep-research reports, and include source citations for transparency. For instance, within the configuration settings for the starting code you can change the length of response from 4,000 words to 8,000 words.
3. Report Presentation: Decide on the formatting of the responses and report generation you prefer with default formatting options like APA, MLA, CMS, Harvard style or IEEE or use Power Point customization startups like Flashdocs.

Best Practices for Secure and Effective Implementation

• Respect user privacy: Utilize OAuth, minimal data permissions, and secure storage.
• Ensure security and compliance: Encrypt data at rest and in transit, clearly communicate data handling practices, and adhere to enterprise security standards.
• Test extensively: Evaluate your system across multiple query types and data sources to identify potential pitfalls using tools like Basalt or Comet to see the inputs and outputs of your AI system and features.
• Utilize Frameworks: Use frameworks like LangChain and LlamaIndex to significantly accelerate development by managing boilerplate AI tasks.

What Can Go Wrong If Not Performed Correctly? Real-World Lessons from Existing Solutions

• Lack of Value: The solution fails to provide meaningful insights or capabilities beyond standard search tools, resulting in poor ROI despite significant implementation costs.
• Internal Data Leak: Insufficient access controls or improper permission settings allow sensitive company information to be accessible to unauthorized employees across departments.
• External Data Leak: Security vulnerabilities in APIs, encryption, or hosting environments lead to company confidential information being exposed to external parties or competitors.
• Spend Too Much Time Setting It Up: Technical challenges, integration issues, and underestimated complexity cause project delays and divert valuable resources from core business activities.

Conclusion

Building a company GPT requires some technical proficiency and a clear understanding of who is using what data when. By creating a system that can integrate multiple data sources securely and effectively it enables a robust AI-driven internal knowledge base that can significantly streamline business operations. After you set up this initial AI use case, your customized GPT will not just provide answers— it will become a strategic asset, accelerating decision-making and driving lasting competitive advantage.

Want to continue the conversation?

Contact NorthLawn to explore how NorthLawn's AI Practice can support your organization's goals.