Building full-firm knowledge graphs to dramatically speed up agents

A firm-wide knowledge graph links matters, documents down to individual clauses, clients, authorities, know-how, correspondence, deadlines and billing so an agent can query structured context instead of re-reading raw files. Most relationships already exist in practice management, registries and ledgers; extraction follows a "rules first" pipeline that parses obvious facts for free and sends only ambiguous sentences to a small LLM. In a test room of 245 documents (8,939 sentences) rules flagged 141 sentences (1.6%) for model reading, costing about $0.047 versus $1.29 to read every sentence. Graphs can be built from JSON exports or from a matter-centric platform; one build produced 6,333 nodes and 7,563 edges in about 1.6 seconds on a laptop with zero model tokens consumed.

Giving an agent the graph produced dramatic speed and cost gains while preserving accuracy. Across five firm-wide questions the agent ran 73% faster, read 92% fewer input tokens and made 66% fewer model calls (total time 509s → 139s, tokens 1.35M → 102K, calls 205 → 30), with accuracy improving to 100% overall; average run cost fell from ~$0.17 to ~$0.02. The hybrid rules+small-model approach was far cheaper and often more reliable than sending every sentence to a model. The test used clean, redrafted text that preserved facts, so scanned or messier real-world inputs will require further handling.