Devlog · 2026-09-24

Engineering Hybrid RAG: When Semantic Vectors Meet Knowledge Graph Entities

Relying solely on vector embeddings overlooks precise entity definitions, while simple keyword matching lacks semantic intuition. Here is how Accrue Notes implements hybrid graph-vector scoring on mobile devices, and why white-box AI reasoning is non-negotiable.

Key Takeaway

Black-box chatbots are the antithesis of rigorous thinking. In knowledge inquiries, we refuse to accept generative hallucinations that pretend to know the answer. Accrue Notes employs a custom Hybrid RAG (Semantic Vectors + Entity Graph Topology) pipeline, exposing the complete AI thinking process and source citation quotes directly to the user.

The Pitfalls of Naive Vector Search

Most conventional AI note tools simply run generic vector embeddings across text chunks and dump the top results into an LLM context window:

Entity Blindness: When notes link [[Occam's Razor]] with [[First Principles]], vector search often retrieves paragraphs sharing superficial vocabulary while missing the exact canonical definition;
Lack of Logical Auditability: When asked a strategic multi-hop question, standard models return plausible-sounding generalities. Users cannot verify which specific notes were referenced, making hallucinations dangerous in decision-making.

Our Approach: Dual-Engine Hybrid Retrieval

[User Inquiry / Research Report Topic]
       │
       ├──► Vector Engine: Local embedding cosine similarity recall (Top-K)
       │
       └──► Graph Engine: Extract [[Entities]], traverse 1st & 2nd-degree bi-directional links
               │
               ▼
       [Reranking & Context Filter with Graph-Topology Weighting]
               │
               ▼
       [White-Box Prompt: Enforce <thinking> chain-of-thought & [Ref] quotes]
               │
               ▼
       [Streaming UI: Real-time collapsible thinking pill + verifiable citations]

Why White-Box Thinking Matters

In our latest release, the knowledge QA pipeline makes reasoning steps fully auditable. As the model synthesizes your knowledge base, every deduction step— "First extracting Taleb's definition of payoff convexity from [[Antifragility Notes]]..." "Next cross-referencing inverted arguments in [[Munger Latticework]]..." streams live inside a collapsible [🧠 Explored in Depth] capsule. Once complete, it folds neatly into a clean, authoritative executive summary that you can inspect at any level of depth.

← All posts