You need data from multiple places
Your data is not in one place and might never be. Your AI will have to consume these silo-ed data and business logic as one-off tools.
Instead, if this fragmented reality was presented to your AI application as a unified query tool, performance would go up 10x.
PromptQL instantly gives you a natural language API which writes and executes a Python and SQL-like query on top of structured, unstructured and API data sources to find and retrieve the most relevant data.
You need an authorization model
Current authorization approaches are untenable to securely make your data available to AI applications, especially with fine-tuning, RAG or traditional text-to-SQL.
PromptQL with Hasura DDN provides a granular model level security mechanism on the semantic object model that sits outside of the source.
Semantic Object Model
Authorization policies are created and managed at this model layer. With attributed based policies to control access at a row, column and method level.
You need flexible planning and orchestration
AI changes the way we interact with our systems. We expect the flexibility of natural language upfront.
The PromptQL API has AI that can create plans to reliably and consistently access & operate on data, interleaved with LLM generation based on 3 key ideas:
Restricted to use a standardized SQL-like query language
This creates a 10x improvement in the “reasoning ability” required for autonomous planning because query language semantics are already embedded.
Self-healing python runtime to create multi-step orchestration plans
This enables LLMs to have an “analytical brain” to perform complicated data tasks, call downstream services, auto-correct itself, and perform bulk operations/iterations on data way beyond context windows.
PromptQL creates, saves and retrieves context in structured memory artifacts
This creates a 10x improvement in the “correctness” of plan execution, by eliminating context loss between multiple steps and scope for hallucinations.
You need to reduce fragility with every failure
Each “failure” of your AI application to do the right thing, should result in an action that makes the end to end system better - from user to data.
Current approaches have too many variables, and improving the system increases the overall fragility.
PromptQL provides a query plan for every run, that helps you focus on 2 key aspects to drive improvement:
Improve the prompt given to PromptQL
PromptQL will not always be able to create the right query plans, but just like any LLM, it can easily be nudged and improved in the right direction.
Improve the data connected to Hasura DDN
PromptQL is able to “reason” about creating plans by leveraging the semantic documentation and the underlying data modeling (move between structured & unstructured data, use business methods for specific intents) that you do.