Choose a Knowledge Source
SyntheticBrew supports several ways to give an agent domain context. Choose based on the shape and freshness of the source, not on the name of the technology.
Knowledge or Knowledge Graphs?
Section titled “Knowledge or Knowledge Graphs?”| Requirement | Knowledge | Knowledge Graphs |
|---|---|---|
| Source | Documents and narrative text | Typed entities and relationships |
| Retrieval | Semantic similarity over chunks | Exact, schema-aware entity lookup |
| Best for | Policies, manuals, articles, support content | Catalogs, codes, taxonomies, registries |
| Update | Upload the current document set | Apply a bundle or upsert entities |
| Agent tool | knowledge_search | list_<type>, get_<type>, optional ID listing |
These capabilities are complementary. A support agent can search policy prose in Knowledge while looking up exact plan identifiers in a Knowledge Graph.
Knowledge Graph or external MCP server?
Section titled “Knowledge Graph or external MCP server?”Use a SyntheticBrew Knowledge Graph when:
- the domain has a stable JSON Schema;
- records form a curated, moderate-size catalog;
- exact IDs and typed filters matter;
- you want SyntheticBrew to generate safe agent tools from the schema;
- changes can be reviewed and applied as configuration.
Use an external MCP server when:
- data changes continuously in an operational system;
- the source must remain the system of record;
- queries require custom business logic, transactions, or permissions;
- the dataset or query pattern exceeds the graph limits shown for your deployment.
The hybrid pattern is common: put stable taxonomy data in a graph and expose live inventory or account data through MCP.
When to use a specialized graph system
Section titled “When to use a specialized graph system”Knowledge Graphs provide typed entity lookup, references, filters, and generated tools. They are not a general graph-computation engine. Use a specialized system such as Neo4j when you need multi-hop traversal, shortest paths, centrality, or large-scale relationship analytics. Expose the required operations to the agent through an MCP server rather than asking the model to generate unrestricted database queries.
When extraction-oriented tools fit better
Section titled “When extraction-oriented tools fit better”Third-party open-source libraries such as LlamaIndex and Cognee can extract entities and relationships from unstructured documents. They are useful when extraction is the goal and the domain is not already modeled. SyntheticBrew Knowledge Graphs are declarative: you define the accepted schemas and entities, and the platform validates them before exposing tools.
Decision guide
Section titled “Decision guide”| Goal | Recommended approach |
|---|---|
| Search manuals or policies | Knowledge |
| Retrieve exact products, codes, or classifications | Knowledge Graphs |
| Combine prose with exact catalog data | Both capabilities on one agent |
| Query a live application or database | External MCP server |
| Run graph algorithms | Specialized graph system through MCP |
| Extract a graph from documents | An extraction pipeline, then validate before import |