If you ask a frontier model a question, you’ll likely get a strong, detailed answer — but it’s probably the same one your top competitor’s partner will get when she asks the same question.
Ask Intapp Celeste, on the other hand, and you’ll get a robust, defensible response built on your firm’s own deals, decisions, and institutional judgment — compounding alpha instead of leveling it.
That’s the semantic layer at work: Before every response, Celeste gathers the context that makes an answer relevant to you, your industry, and your firm, so you never have to re-explain that context yourself. Celeste assembles that context automatically on every turn and learns as you provide additional information.
Two layers underpinning Celeste’s semantic layer
Celeste delivers tailored answers through two layers: the connector level (what each connected system contains) and the firm model (how your firm thinks about its business).
At the connector level, Celeste keeps a semantic model of each connected system, mapped one-to-one to its entities and fields. Because mapping is direct, Celeste builds it automatically, so teams don’t have to map anything by hand.
Administrators can then add what a raw schema can’t express: metrics, rules, concepts, and query examples for that specific system, like a proven query pattern for a tricky question. This layer goes beyond the open standard (OSI, see below), so admins create it manually for each connector if desired.
Above the individual systems sits the firm model: entities, relationships, metrics, rules, and concepts, all in your firm’s language and independent of where the data lives. Admins curate the firm model, which Celeste initially populates from a template.
“Source mappings” connect the firm model to the underlying connector models, mapping the firm’s entity properties to the corresponding objects and fields in each connector.
When you ask Celeste a question, it uses an embeddings model to check both models and find the chunks of context that match what you’re asking. The model then gives that context to Celeste to inform the response.
How the semantic layer is built
The semantic layer is built on an open standard so context can travel between tools.
Celeste’s semantic models follow the Open Semantic Interchange (OSI), an industry standard for describing what business data means. OSI enables:
- Portability: A semantic definition written for OSI is never proprietary to Celeste. If your firm has semantic models in another OSI-aligned tool, that work comes with you.
- Common language: Celeste organizes context using the same building blocks the market standardizes on, not a custom format only Celeste understands.
- Future-proofing: As more platforms adopt OSI, the information you capture once becomes reusable across the systems you already own.
Seven foundational building blocks make up the model of how your firm works:
- Entities: The terms that effectively describe your business, such as deal, fund, client, and matter. Ask for the latest team update on a client, and Celeste finds the correct update record instead of guessing from a meeting note.
- Properties: Details describing each entity. Each deal has a value, a close date, a stage, and each fund has a vintage year.
- Relationships: How entities connect. Celeste follows the natural links in your business. Each deal belongs to a fund; a matter is opened for a client. Celeste uses these relationships to understand and provide the full context.
- Metrics: Calculated figures that your firm reports. Define the metric once and Celeste computes it consistently every time. For metrics with multiple valid formulas, such as total deal value or IRR, Celeste can store each definition and use the one you specify.
- Rules: Constraints your firm holds true. These rules keep answers consistent with how your firm operates instead of experimenting with what the data technically allows. Every deal must have a responsible partner, and a commitment can’t exceed its fund’s target.
- Concepts: Your firm’s vocabulary. Celeste knows your firm’s definitions of terms like “committed capital,” “churn,” and “engagement,” along with the synonyms people use for them.
- Source mappings: Links between the model and your data. These links tell Celeste where to find each entity; for example, a deal may live in a specific table in your warehouse or a set of fields in your CRM.
Schemas constantly change. A warehouse may get restructured or a CRM field renamed, and your semantic model has to keep pace. Regenerating a connector’s model refreshes its entities and relationships to match the new schema — without touching the rules, metrics, and concepts manually added by an admin.
If the updated schema removed a dependency for a rule or concept, that rule can become stale. Celeste automatically surfaces this issue and proposes its own updates, so the semantic layer continuously improves.
Key takeaways
- Responses are consistent with your firm’s language and context: Define a term like “committed capital” once, then Celeste will consistently apply that definition across firmwide, keeping your teams aligned.
- Context is never locked into Celeste: Keep prior work portable with OSI-based models, so any work you do in Celeste isn’t stranded if you adopt another OSI-aligned tool later.
- Complex answers get faster and more reliable: Get better responses — not guesses — on the first try with Celeste, which understands the intent behind your question.
- Anyone can get a clear answer: Translate your firm’s vocabulary and constraints into something Celeste can understand and query, so users don’t have to know where the data lives or what it’s called in your systems.
See Celeste in action — request a demo.