How it works
The agent is grounded in a knowledge graph that lives in its own GitHub repo and evolves daily via scheduled Claude routines. Every claim the agent makes cites a specific node ID in that graph — you can click the citation to read the underlying source.
The three repos
cdp-recommendation-agent-kg — the knowledge graph itself. Markdown-with-YAML files, one per node, with a JSONL edge file. Schema docs define node types (Concept, Constraint, Pattern, Tradeoff, Vendor, …) and edge types (alternative-to, governs, applies-to-domain, …).
cdp-recommendation-agent — the routines that reflect on conversations and evolve the graph. Eight scheduled routines: Web Knowledge Refresh, Technical Reviewer, Organization Reviewer, Cross-cutting Synthesizer, Question Catalog Evolver, Code Change Proposer, PR Author, Marketing Content Drafter.
cdp-recommendation-agent-ui — this app. Reads the KG at build time, runs the chat at request time, persists transcripts to feed the next day's routines.
Bias-neutrality enforcement
- Every Vendor node in the KG must have at least one
alternative-toedge to another Vendor. - Every Tradeoff node must define both ends with associated nodes.
- The Cross-cutting Synthesizer routine audits these constraints daily and flags violations.
- The agent's system prompt enforces "every Vendor mention paired with at least one alternative" as an output rule.
Currently grounded on
Live counts of the KG vocabulary the agent reasons over:
- Concepts0
- Vendor Constraints0
- Tradeoff Axes0
- Strategic Patterns0
- Vendors0
- Latency Tiers0
- Modalities0
- Use Cases0
Want to start a structured conversation? Take the four-section assessment first — the agent will tailor its recommendations to your specific profile.
Start the assessment →