Our tech stack is not a simple “wrapper” over an LLM or an app like a GPT. Instead, it’s an implementation of Retrieval Augmented Generation (RAG) that puts individual expertise at the center, and leverages an API to reference an LLM only when useful.
In testing, a RAG approach has been proven to dramatically reduce the infamous “hallucinations” that LLMs can produce. In addition, it makes the answers specific to you making them differentiated vs. general.
With a RAG approach, that content that you upload with Tmpt, as well as your profile, preferences, and so on, are stored in a special AI-enabling vector database that is referenced first when an answer is being generated. We call it your Tmpt Knowledge Base (KB).
The content you add is “vectorized” by splitting into individual passages that can be accessed and assembled in new, generative ways with other passages. The vectorization process captures the semantic content of each passage so that questions can be answered with relevant passages, even if the wording is different (potentially even in a different language!)
Google Cloud is trusted by some of the most secure companies on the planet, and maybe even by you (for Gmail, Photos, and other sensitive personal information). In their own words: “Google provides the world's most advanced security that keeps more people safe online than anyone else. Our approach, developed over decades operating at a global scale, comprehensively protects your data. See their approach here.
We don’t leverage your knowledge base for the benefit of anyone but you and your seekers. We would never sell any of your data to anyone.
Other people can’t see or access your content. All information in your Tmpt KB is private to you and the people you invite; no outsiders ever have access, unless you invite them. Even then, we put you in control with preference settings such as whether or not sourced content is viewable.
Like most major online services, Tmpt personnel will, on rare occasions, have access to your content but only
1) when legally required
2) when necessary to ensure our services are working properly or to inform service improvements
3) when enforcing our Terms of Service and Privacy Policy.
Access is limited to a small number of people via a strict policy and technical access controls that prohibit access to file content except in these rare circumstances. In addition, we use a number of physical and electronic security measures to protect user information from unauthorized access.
Once the Tmpt service has searched your KB for relevant passages from your sources to answer a question in a way that’s uniquely you, we may interact with an LLM to add natural language, context, and other elements that make the answer especially impactful.
While the companies that create LLMs do crawl the open web for information, the API terms of the LLMs we use stipulate that they do not train on the data we send them through their API.
From OpenAI: “We do not train on your business data” (https://openai.com/enterprise-privacy/).
From Google: “Google doesn't use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products” (https://ai.google.dev/gemini-api/terms).
From Anthropic: “We do not use API prompts or conversations for any model training purposes” (https://www.anthropic.com/uk-government-internal-ai-safety-policy-response/data-input-controls-and-audit)
With Tmpt, you can always edit an answer your AI Assistant has generated on your behalf, and if you choose you can do so before it’s ever shared with anyone. This gives you the option of full control over the curation of any output leveraging your KB content that would be accessed by anyone in the world.
When you take the time to revise an AI-generated answer, give it a glowing rating, or create your own, the Tmpt system captures that answer as your own “gold standard” response to that question, so that over time the system answers in the way you prefer.
We’ve made it easy to redact portions of any conversation transcript, delete any content you’ve uploaded, or even delete all conversations or your entire KB, all through your dashboard.
Of course, we’ll ask “Are you sure?” so you don’t ever do it by accident.
We love to share stories of how our service helps Experts help others, but we will never do that without your permission.