Knowledge Graph
Every SAP landscape is unique, shaped by years of custom development and configuration decisions.
Adri builds a Knowledge Graph of that unique landscape to make AI agents work for you.
When a customer connects their SAP system with Adri, a dedicated Knowledge Graph is built. The Knowledge Graph is a structured representation of every custom object and their dependencies in a customer's SAP system.
This gives Adri agents deep, system-level context. So when a user asks Adri agents to research, code or document, the agents understand your custom tables, your enhancements, and your configuration at scale.
How large is this Knowledge Graph?
For a typical S/4HANA 2023 system with 100,000 custom objects, the Knowledge Graph will contain about 30 million records. It takes about 250GB to store this in a vector database.
Try the free demo sandbox to see large-scale knowledge creation and retrieval in action. The Knowledge Graph in this sandbox contains about 27 million objects for S/4HANA 2023 and 18 million objects for ECC 6.0.
How It's Built
The Knowledge Graph is constructed entirely within the customer's VPC through the following pipeline:
-
Extraction — The Adri AI Connector, running in the SAP-restricted subnet, reads ABAP objects from SAP Gateway via the standard ADT API.
-
Outbound push — The Connector pushes the extracted objects to the Knowledge Graph Builder in the customer's VPC over a Secure WebSocket connection (mTLS + JWT). The SAP subnet's "no inbound ports" policy is preserved because the Connector initiates all connections outbound.
-
Embedding — The Knowledge Graph Builder sends ABAP content to an LLM Embedding Service via VPC endpoints (PrivateLink). The content sent is determined by the customer-configured content scope — full source, signatures and documentation only, or local embedding only. Embedding traffic never traverses the public internet.
-
Storage — The Builder distributes the results across three storage tiers based on data size and query requirements.
The customer controls what ABAP content is sent for embedding. Three levels are available: full source (default), signatures and documentation only (method bodies stripped), or local embedding only (no data sent to any LLM provider). See LLM Embedding Service for details.
How It's Stored
The Knowledge Graph is stored across three tiers, all within the customer's VPC:
| Storage Tier | What It Holds | Size Limit | Purpose |
|---|---|---|---|
| Vector Database | Embedding vectors + minimal metadata (object name, type, package) | 8,096 bytes per record | Semantic search — find objects by meaning |
| Postgres Database | Source code, dependency relationships, structured metadata, documentation | 1 MB per record | Keyword search, full-text search, relational queries |
| S3 Object Storage | Oversized ABAP objects (large reports, class pools, generated code) | Effectively unlimited | Objects exceeding 1 MB stored without truncation |
This tiered approach ensures no single storage system is forced to handle data outside its optimal range. The Vector Database and Postgres store references (S3 keys) to oversized objects, so the query engine can still locate and fetch them on demand.
How It's Queried
The Adri AI Server handles all queries against the Knowledge Graph. It authenticates developers via the customer's IdP, enforces authorization policies, and then executes queries through its embedded ChromaSQL engine.
ChromaSQL supports three search methods:
- Semantic search — queries the Vector Database to find objects by meaning (e.g., "special stocks from vendor")
- Keyword search — queries Postgres for exact matches,
LIKEpatterns, and structured filters (by package, object type, transport request) - Large object retrieval — fetches oversized objects from S3 when needed
Results from all three tiers are merged and ranked before being returned to the Adri server.
What Stays in the VPC
All Knowledge Graph data — embeddings, source code, metadata, dependencies, and oversized objects — is stored within the customer's VPC. No Knowledge Graph data leaves the VPC. The only data that crosses the VPC boundary is the content sent to the LLM provider for embedding, and the customer controls the scope of that content.
The Knowledge Graph Builder currently supports ECC and S/4HANA systems. Support for additional SAP products (SuccessFactors, etc.) and add-ons (OpenText, etc.) is coming. This will extend the Knowledge Graph beyond the core ECC and S/4HANA landscape to cover the broader SAP ecosystem.
FAQs
Can I connect multiple SAP systems to Adri?
Yes. You can connect any number of SAP systems — irrespective of version (ECC, S/4HANA, etc.). A dedicated Knowledge Graph is built for each connected system, so agents have system-specific context for every landscape in your environment.
Can I control what data is indexed?
Yes. The customer configures a content scope that determines what ABAP content is sent for embedding — full source code, signatures and documentation only, or local embedding only (no data leaves the VPC). This can be set per SAP system. See LLM Embedding Service for details.