Nova: Structural GraphRAG System

"Don't just find facts—find roles."

Project Nova is an advanced Retrieval-Augmented Generation (RAG) system that bridges the gap between Semantic Search (content) and Topological Reasoning (structure). Unlike traditional RAG, which relies solely on text similarity, Nova uses a Knowledge Graph to "route" queries to structurally relevant context.

---

🏗 Architecture: "Graph as Recommender"

We have moved beyond simple text-chunking to a Dual-Store / Router Architecture.

1. The Core Loop

The system does not inject raw graph triplets into the prompt (which confuses LLMs). Instead, it uses the graph to expand the search space:

graph LR
    Q[Query] --> T[Topic Extraction]
    T -->|Anchor Entity| G[Graph Store]
    G -->|Node2Vec| S[Similar Structural Nodes]
    S -->|Entity Lookup| V[Vector Store]
    V -->|Original Text| C[Structural Context]
    Q -->|Embedding| V2[Vector Store]
    V2 -->|Cosine Sim| C2[Semantic Context]
    C --> P[LLM Prompt]
    C2 --> P

Loading

2. The Stores

• Vector Store (v_store): Holds the specific, rich text chunks.
• Triplet Store (t_store): A "Card Catalog" for the graph. Maps entities to triplets.
• Graph Store (g_store): A NetworkX graph purely for topology (connectivity).
• Node2Vec Model: A trained embedding model that understands graph structure (roles) rather than text meaning.

3. Dual-Context Prompting

The LLM receives two distinct streams of information:

• SEMANTIC CONTEXT: "Here are facts directly related to your keywords." (e.g., Cat sleep times)
• STRUCTURAL CONTEXT: "Here are things that play a similar role in the system." (e.g., Other predators that hunt in this specific way)

---

🧪 Stress Test Results: cat_facts.txt

We ran extensive stress tests to validate the "Structural Analogies" capability.

Question	Type	Nova's Performance	Key Insight
"Who invented the cat flap?	Factual	✅ Perfect (Isaac Newton, Spithead)	Semantic retrieval handled this easily.
"Why must a cat back down a tree?"	Structural	✅ Best Answer	The graph correctly linked "claws" to "climbing mechanics" without direct keyword overlap.
"Felicette of Egypt?"	Analogy	❌ Hallucination	LLM hallucinated "Daenerys" due to training data leakage. Fix: Stricter prompt instructions.
"Jon Snow in Answer"	Bug	❌ Contamination	Stale graph data from a previous run persisted. Fix: Implemented strict rm *.pkl hygiene.

Top Finding: The "Graph as Recommender" pattern eliminated the "broken English triplet" problem, resulting in much higher quality answers than raw graph injection.

---

🚀 Research & Roadmap

Current Feature: Node2Vec (Active)

• What it does: Learns low-dimensional representations of nodes based on random walks.
• Why: It captures "Structural Equivalence". Nodes that bridge communities look similar, even if they don't share neighbors.

Next Step: Hyperbolic Embeddings (Planned)

• Problem: Euclidean space (dot product) creates distortion for hierarchical trees (taxonomies).
• Solution: Migrate the Graph Store to a Poincaré Ball model.
• Projected Gain: Better representation of "Is-A" hierarchies with fewer dimensions.

Production Comparison

Feature	Project Nova	Microsoft GraphRAG
Structural Analogies	✅ (Node2Vec)	❌
Ingestion Quality	⚠️ (Needs Validation)	✅ (Entity Resolution)
Retrieval Mode	✅ Dual-Context	✅ Community Summaries