Since we launched the Delphi dashboard, the most consistent feedback from clients has been some version of the same request: "I don't want to build filters and scan tables. I just want to ask a question and get an answer."
That's fair. Not every intelligence need starts with a structured query. Sometimes you want to know "what's the status of hydrogen projects in the Netherlands that use PEM electrolysis?" — and you want the answer in a sentence, not a pivot table.
That's why we're building Aletheia.
Aletheia is an AI-powered question-answering interface built on top of our knowledge graph. Ask a market intelligence question in plain language, get a structured, source-backed answer. No hallucinations, no guesswork — just the data we actually have, presented clearly.
It's not ready yet. But the architecture is far enough along that we want to share what we're building and why we think it matters.
The architecture: Graph-RAG, not vanilla RAG
Most AI-powered market intelligence tools work by feeding documents into a vector database and using retrieval-augmented generation to find relevant chunks of text. This works reasonably well for general Q&A, but it has fundamental limitations for structured market intelligence.
Vector search finds text that's semantically similar to your question. It doesn't understand relationships. It can't tell you that a membrane supplier connects to an electrolyzer manufacturer that connects to a hydrogen project that connects to an offtake partner — because those connections aren't in the text chunks. They're in the structure of the market itself.
Aletheia uses what we call Graph-RAG: retrieval-augmented generation where the retrieval layer queries a knowledge graph, not a vector store. When you ask a question, the system parses it into a graph query — identifying entities, attributes, and relationships — and traverses our structured data to find the answer.
A question about "hydrogen projects in the Netherlands using PEM" gets parsed into: find nodes of type "hydrogen project," filter by geography "Netherlands," filter by technology attribute "PEM electrolysis," return matching entities with their key attributes. The graph query returns structured results with full provenance. A language model then narrates those results into a natural-language response.
Why this matters for trust
This architecture means two things that matter for enterprise use.
Every factual claim in the response traces back to structured data in our graph. There's no "the AI read some articles and guessed." When Aletheia says a project is in development phase, that status comes from a specific data point with a source and a timestamp. You can click through to see it.
The response is bounded by what our data actually contains. If we don't have data on a topic, the system says so — it doesn't improvise. For casual research, that constraint might feel limiting. For decisions involving capital allocation, procurement strategy, or policy design, it's exactly what you want. Knowing what isn't in the data is as important as knowing what is.
What you'll be able to ask
We're designing Aletheia for questions that map to our knowledge graph. Some examples across sectors:
"Which companies are developing direct air capture projects in Europe, and what's the combined announced capacity?" — traverses our CCUS graph to return DAC developers with project details and sources.
"What electrolyzer manufacturers have supply agreements with projects in the Middle East?" — follows supply chain edges in the graph to find manufacturer-to-project relationships in the specified region.
"Compare the hydrogen project pipeline in Germany and Spain by project phase." — aggregates project counts by status for each country and returns a structured comparison.
Follow-up questions will work within the same conversation. "Now show me only the ones that have reached FID" narrows the previous result set. "Which of those use alkaline technology?" filters further. Context carries forward so you can drill into the data iteratively.
What it won't do
Aletheia won't generate market forecasts, investment recommendations, or strategic advice. It won't summarize news articles or provide commentary on policy developments. It answers factual questions about the structured data in our knowledge graph.
If you ask something outside our data coverage — "what's the best investment in renewable energy right now?" — the system will tell you that the question is outside its scope, rather than generating a speculative answer.
We think this constraint is a strength. There are plenty of tools that will give you fluent but unreliable answers to broad questions. We'd rather build something that gives precise, auditable answers to specific ones.
Timeline
Aletheia is in internal testing now. We're working through the edge cases — ambiguous queries, multi-hop graph traversals, response quality for questions that span multiple sectors — before opening it up more broadly.
If you're a current Delphi platform user and want early access when it's ready, let us know. And if you have questions about the architecture or want to discuss how grounded AI fits into your intelligence workflow, we're always happy to talk.