hila is a research platform with generative UX chat capabilities that provides accurate answers to questions about public companies. The world is quickly becoming more familiar with chatbot technology, and hila is a practical use case of Large Language Models (LMMs) pointed toward a specific subset of public data like earnings calls, SEC filings, and 10Ks.
Read more about how hila specifically works here.
A key feature of how hila produces its answers — which are incredibly meticulous due to patent-pending technology leading to a wildly low hallucination rate — is prompt engineering.
What is Prompt Engineering?
Prompt engineering is a technique with a largely self-explanatory name used in natural language processing (NLP). Engineers feed language models text prompts, called inputs, and then rate the model’s answers, called outputs. This process is a way for the humans behind the technology to lead and monitor AI, a pillar of hila and our parent company Vianai. Human-centered AI keeps the technology honest, useful, and safe, a key aspect of hila’s prompt engineering technique.
Prompt engineering is essential for generative NLP tasks such as text classification, language generation, and answering questions. The engineers in charge of this process must ensure that:
- The model is trained on accurate and sufficient data
- The model is given multiple formulations of the prompt
- The task provided to the model is clear and contextual
Prompt engineers must anticipate the questions a user might ask and then feed the model input text with enough context to provide useful output generation. They must tell the model what to expect from users in clear, simple language and provide variations and spelling mistakes, even trying different writing styles, voices, and tones to mimic the questions a user might ask.
The engineer can also show the model the desired outputs, called “few-shot learning.” Few-shot learning is a technique by which the model is provided, via the prompt itself, with examples of possible inputs and the respective expected outputs. This helps the model generate outputs in either the format, complexity, or style desired by the end user.
Another key aspect of prompt engineering is to select the right language models for the purpose of the application. Different language models have different strengths and weaknesses, and some may be better suited to certain tasks than others. For example, while GPT 3 to 4 can be used for generating coherent and fluent language, models such asCohere's summarization model may be better suited for summarizing long documents.
hila’s language models
hila uses a variety of high-quality open-source machine-learning models built to provide accurate, reliable, and effective natural language responses to queries. We use a document-centric approach while still leveraging the power of LLMs and cite the sources which our models pulled answers from.
LLMs like GPT3 and GPT4 are prone to hallucinating. Despite how fluent their outputs may seem, they can be inaccurate, entirely made up, or worse - harmful. This is due to how these models are designed. There is no inherent fact-checking module or internal truth center in these models. Rather, they have been trained to take inputs and generate outputs, regardless of their fidelity or accuracy. Steps have been taken by OpenAI and other companies to mitigate erroneous or harmful outputs, but the issue is far from solved.
The team at hila and Vianai have developed state-of-the-art techniques to analyze the outputs of these LLMs to ensure that they aren't erroneous or present harm to an end user. Through the use of both new and old NLP techniques and additional models, hila outputs a proprietary "hila score," which is a measure of how well supported a provided answer is by evidence. In addition, hila determines the exact sentences of text that were used to provide a given answer. Users of the hila platform can expect to not only get answers generated by LLMs but also measures of confidence and direct links to sources for answers.
hila is unlike other research platforms harnessing the power of generative AI. Our intuitive platform can cut down research time from hours to minutes while providing accurate, secure information.
Get started with hila for free today and see the wonderful results of our expert prompt engineering!