Reexpress Connect preview.

Building AI co-pilots/agents/teammates has been a non-starter for enterprise settings with existing inference methods and models because the lack of uncertainty quantification compounds with each step of a pipeline, rendering the whole system brittle, unreliable, and ultimately, unusable.

We have a solution.

Reexpress Connect extends the capabilities of our on-device application to a cloud endpoint, providing reliable, robust, interpretable, and fast (!) uncertainty estimates (a.k.a., statistically robust guardrails) for essentially any neural network deployment.

Reexpress Connect is important both for processes that happen behind the scenes from the user (e.g., for building agent pipelines that can re-ask/re-prompt automatically and seek outside tools when needed), and also critically for the end user, as well, so that they know when they can trust the output as-is vs. when they need to take a closer look.

Reexpress Re-search: Search with confidence

Reexpress Re-search is a demo illustrating how Reexpress Connect fits into a generative AI pipeline with retrieval, in this case where the retrieval is web search, providing the missing layer of reliability for generative AI models: reexpress.io

The model is dynamically re-asked/re-prompted if the initial response is estimated to be unreliable. The demo only executes this iterative re-ask process a max of one time, and the responses are constrained to be very concise, but enterprise versions can continue that process to increased depths, with longer responses, over your own document databases.

Corral LLM hallucinations

For retrieval, adding Reexpress to your system has the value of dramatically reducing unexpected surprises from hallucinations and highly confident wrong answers. Out-of-distribution inputs and outputs can be reliably rejected, saving users time and avoiding introducing subtle errors into downstream decision-making process.

For example, it has been pointed out recently on social media that GPT-style models struggle with counting letters in words, e.g., "how many q's in dialogue". These are simple but illustrative examples: The underlying generative model we use in the demo also struggles with these types of questions and will tend to produce incorrect results in the generated text. Nonetheless, the Reexpress constraints are able to correctly indicate that the model is very unconfident about the responses (see image). In practice, this then enables building agent/retrieval pipelines that can short-circuit and take additional action, rather than continuing a series of subsequent tasks using the incorrect results.

Importantly, our uncertainty models in the demo have not been trained with examples similar to "how many q's in X". That would be a major problem for a typical post-hoc output filter because the filter itself would be unreliable; that is, the filter itself could be easily thrown off the rails. In contrast, since Reexpress tightly controls the epistemic uncertainty, out-of-distribution (or otherwise very low reliability) examples such as these will be detected as such.

LLM systems can seem uncontrollable even to experts. A variety of approaches, such as retrieval, have been suggested to improve reliability; however, for anyone using such systems, it quickly becomes clear that something fundamental is still missing. Reexpress is that missing key for building reliable AI systems, providing a novel, robust approach for addressing verification and calibration over neural networks.

Note: In the demo, for reference purposes, we show the text of uncertain outputs, and we also allow the edge case of producing output when there are no relevant retrieved documents (i.e., none of the retrieved documents become highlighted in green). In an enterprise RAG setting, typically the system would be structured to just altogether throw an error in those cases, and not show any output.