Awesome Search KG

Embabel DICE

2 min read

Embabel DICE

The reference implementation of DICE, and a module of the Embabel agent framework — a Kotlin/Spring library for knowledge-graph construction and reasoning over LLM-extracted propositions. Carries an Incubating badge; pre-release.

What It Does

Turns unstructured text into typed, scored propositions, resolves the entities they mention to canonical ids, reconciles them against what is already stored, and projects the result into whichever backend the application needs. Propositions are the system of record; graphs, Prolog facts and agent memory are all projections of them.

The four stages:

  1. Extraction — LlmPropositionExtractor pulls typed propositions with confidence, importance and decay scores.
  2. Entity resolution — mentions resolve to canonical entity ids through an escalating chain.
  3. Revision — LlmPropositionReviser classifies each new proposition as IDENTICAL, SIMILAR, CONTRADICTORY, GENERALIZES or UNRELATED; similar ones merge, conflicting ones lose confidence.
  4. Projection — into Neo4j graphs, Prolog facts, or agent memory.

PropositionPipeline orchestrates the first three.

Entity Resolution

The design worth stealing: EscalatingEntityResolver runs CandidateSearchers cheapest first — by id, exact name, normalized name, partial name, fuzzy name, vector, then agentic — stopping as soon as one resolves. Where several candidates survive, LlmCandidateBakeoff asks a model to choose.

This keeps the expensive paths (vector search, an LLM call) off the common case, which is the same instinct behind tiering models in Agentic Memory for Search Personalization.

Memory Model

MemoryProjection separates propositions by KnowledgeType — SEMANTIC, EPISODIC, PROCEDURAL, WORKING. The Memory class implements LlmReference with both eager and on-demand retrieval, and MemoryMaintenanceOrchestrator consolidates, abstracts and retires memories over time.

ContextId is the primary scoping mechanism for every proposition query: one user may hold several contexts (personal, team, project), and one context may be shared between users.

Stack

Kotlin with Java interop · Spring Boot · Maven · tuProlog for inference · Neo4j for graph projection · Jinja templates for prompts · Apache Tika for file extraction · Micrometer for metrics.

A REST surface exposes /api/v1/contexts/{contextId}/ for extraction, proposition CRUD and search, and entity-scoped memory retrieval, with optional X-API-Key authentication.

  • django-dice — an independent Python/Django implementation of the same design

Graph View

Backlinks