Knowledge Graph RAG (GraphRAG) Uplift Plan: 2025 Roadmap

Status (2026-10-13): A historical planning document; the plan was not executed as written, and the target dates in section 5 passed without those milestones. What happened instead:

• Hybrid retrieval shipped as BM25 + dense with Reciprocal Rank Fusion (HybridRetrieval), and RAG context is assembled by the version-pinned, cited context bundle (GET /api/bundle, assemble_bundle) rather than a graph-plus-vector fusion.
• "Path B" (Knowledge Graph traversal feeding ranking) was never built as the parallel retrieval path described below. The nearest thing that was built, a Knowledge Graph rerank stage over the fused results, measured zero net lift in a 2026-06-16 ceiling experiment even with a Claude-quality Knowledge Graph, and was deleted in 2026-07. See KnowledgeGraphRerank. The accuracy targets in section 1 were never validated against this corpus.
• LLM triplet extraction shipped as the Ollama chunk-level and page-level entity extractors (default gemma4-graph:12b), feeding a proposal queue with an asynchronous judge and human review (WikantikKnowledgeGraph, KnowledgeGraphExtractionBenchmarks). canonical_id stability shipped with the structural spine (StructuralSpineDesign).
• The trust tiers became a machine / human review tier on every node and edge (ProvenanceModel); human-vetted and ai-reviewed as used below are not the current terms.

As of mid-2025, pure vector-based Retrieval Augmented Generation (RAG) has reached a performance ceiling. While effective for simple semantic recall, it fails spectacularly on schema-bound and multi-hop reasoning queries. This document outlines the uplift plan to transition Wikantik to a Hybrid GraphRAG architecture.

1. The Performance Gap (2025 Benchmarks)

Recent evaluations using the Retrieval Experiment Harness confirm the industry-wide trend: Vector RAG accuracy collapses as query complexity increases.

Query Complexity	Vector RAG	GraphRAG (Target)	Delta
Simple Factoid	92%	88%	-4%
Multi-Hop Reasoning	22%	78%	+56%
Global Theme Synthesis	5%	85%	+80%
Schema-Bound (e.g. KPIs)	0%	92%	+92%

The "Graph Tax" (latency) is estimated at ~200ms per request, which is acceptable given the massive gains in precision.

2. Phase I: Extraction & Indexing Uplift

The primary bottleneck is currently the quality of entity extraction. We will move from a "Chunk-First" to an "Entity-First" indexing model.

1. LLM-Augmented Triplets: Upgrade from basic regex extraction to gemma4-based triplet extraction (Subject \xrightarrow{predicate} Object).
2. Canonical ID Stability: Ensure every extracted entity maps to a stable canonical_id to prevent graph fragmentation.
3. Benchmark Integration: Use Knowledge Graph Extraction Benchmarks to measure triplet recall on every commit.

3. Phase II: Hybrid Retrieval Architecture

The new pipeline will execute two parallel retrieval paths:

• Path A (Recall): Standard BM25 + Dense Vector search on the chunk_embeddings table.
• Path B (Precision): BFS-based traversal of the kg_edges table starting from identified entities in the query.

The results are fused using Reciprocal Rank Fusion (RRF) before being presented to the LLM context window. This architecture is detailed in the Page Graph vs Knowledge Graph comparison.

4. Phase III: Validation & Quality Control

To prevent regressions, the following controls will be implemented:

• MCP Guardrails: All graph mutations must pass the McpIntegrationTestFix suite.
• Trust Tiers: Only human-vetted or high-confidence ai-reviewed nodes will be admitted to the Path B retrieval loop.

5. Timeline

Milestone	Target Date	Status
Baseline Benchmark	2026-04-24	[DONE]
Triplet Extractor Upgrade	2026-06-01	[PLANNED]
Hybrid Fusion Engine	2026-07-15	[PLANNED]
Production Cutover	2026-08-30	[PLANNED]

This plan is managed by the Infrastructure & SRE team.

---

Cluster Hub: Wikantik Platform Development Hub: Architecture, MCP Protocols, and Agent Briefings (wikantik-development)