Making the Knowledge Layer a Graph You Actually Traverse
The author retires the Part 1 query router (wiki-first, evidence-first hybrid decision) in favor of a retriever-versus-filter principle that enables real multi-hop traversal, time-aware relationships, and discovered contradictions Entity resolution fragmentation was a critical failure mode: live extraction produced 149 concept objects where the curated graph had only 19, revealing a ~7x fragmentation factor that demanded embedding similarity and LLM adjudication The new architecture introduces b
Analysis
TL;DR
- The author retires the Part 1 query router (wiki-first, evidence-first hybrid decision) in favor of a retriever-versus-filter principle that enables real multi-hop traversal, time-aware relationships, and discovered contradictions
- Entity resolution fragmentation was a critical failure mode: live extraction produced 149 concept objects where the curated graph had only 19, revealing a ~7x fragmentation factor that demanded embedding similarity and LLM adjudication
- The new architecture introduces bitemporal edges (storing both valid-time and transaction-time), ingest-time contradiction detection, and an always-fused retrieval pipeline that eliminates routing quality dependence on question phrasing
- The implementation is deployed on Azure using Cosmos DB for Apache Gremlin as a second graph engine alongside the existing Azure AI Foundry stack
- The system is validated against a synthetic insurance corpus (Ostermere Mutual), demonstrating multi-hop reasoning with citations at every step and the ability to answer questions the Part 1 architecture could not
Why It Matters
This work addresses two of the most persistent failure modes in production RAG systems: entity fragmentation that degrades knowledge graph quality, and retrieval brittleness caused by query-dependent routing logic. For AI practitioners building enterprise knowledge systems, the shift from router-based retrieval to always-fused graph traversal with bitemporal edges represents a meaningful architectural advancement that improves both accuracy and robustness.
Technical Details
- Retriever vs. Filter principle: Replaces the query router with a design where the knowledge graph is actively traversed rather than consulted like a filing cabinet, enabling multi-hop reasoning chains with citations at every step
- Bitemporal edges: Relationships in the knowledge graph encode both valid-time (when a fact was true in the domain) and transaction-time (when the system recorded the fact), enabling time-aware scoping such as preventing a March 1 rule from applying to a February 20 loss
- Ingest-time contradiction detection: Contradictions are discovered during ingestion rather than curated post-hoc, with a contradiction register that refuses to answer questions the source data has not resolved
- Entity resolution subsystem: Addresses the 7x fragmentation problem through embedding similarity matching and LLM adjudication, forming the foundational layer the entire graph depends on
- Azure deployment: Uses Cosmos DB for Apache Gremlin as a new graph engine alongside the existing Azure AI Foundry stack, with a projection worker and ontology layer for reranking the fused retrieval union
Industry Insight
- The 7x entity fragmentation factor is likely a common but underreported problem in production RAG systems; organizations should audit their entity resolution pipelines and invest in embedding-based deduplication with LLM adjudication rather than relying on alias-based matching alone
- Query-dependent retrieval quality is a silent reliability risk: systems whose performance varies with question phrasing will produce inconsistent results in production, making the always-fused retrieval approach a safer default for enterprise deployments
- Bitemporal graph modeling is particularly valuable in regulated industries (insurance, healthcare, finance) where the distinction between when a fact was true and when it was recorded is operationally critical; this pattern should be considered for any domain with temporal policy or compliance constraints
Disclaimer: The above content is generated by AI and is for reference only.