Building Deterministic Graph-RAG Systems Beyond Vector Search

Beyond Vector Search: Building a Deterministic 3-Tiered Graph-RAG System

This architecture implements a three-tier retrieval hierarchy using a Python QuadStore and ChromaDB to ensure factual integrity. By enforcing a Subject-Predicate-Object plus Context (SPOC) format, the system resolves data conflicts through prompt-enforced fusion rules.

Why This Matters

Standard vector databases are inherently “lossy” regarding atomic facts and strict entity relationships, often confusing specific details like team affiliations due to their proximity in latent space. By implementing a federated architecture that prioritizes deterministic graph data over fuzzy semantic embeddings, engineers can eliminate relationship hallucinations and ensure absolute predictability for critical ground-truth facts.

Key Insights

• 3-Tier Hierarchy (2026): Organizes data into Priority 1 (absolute graph facts), Priority 2 (statistical data), and Priority 3 (vector fallback) to resolve information conflicts.
• SPOC Format: Uses a Subject-Predicate-Object plus Context (SPOC) schema in a QuadStore to enable constant-time lookups across multiple dimensions.
• Prompt-Enforced Fusion: A conflict resolution strategy that embeds adjudicator rules directly into the system prompt instead of using complex algorithmic routing like Reciprocal Rank Fusion.
• Named Entity Recognition (spaCy): Bridges the gap between semantic queries and deterministic graphs by extracting entities to perform strict lookups in the QuadStore.
• Lightweight QuadStore: A Python-based in-memory knowledge graph that maps strings to integer IDs to prevent memory bloat while maintaining high-speed indexing.

Working Examples

Initializing the Priority 1 QuadStore and populating it with SPOC-formatted factual data.

from quadstore import QuadStore

# Initialize facts quadstore
facts_qs = QuadStore()

# Natively add facts (Subject, Predicate, Object, Context)
facts_qs.add("LeBron James", "likes", "coconut milk", "NBA_trivia")
facts_qs.add("LeBron James", "played_for", "Ottawa Beavers", "NBA_2023_regular_season")
facts_qs.add("Ottawa Beavers", "based_in", "downtown Ottawa", "NBA_trivia")

Setting up the Priority 3 vector database using ChromaDB for long-tail context.

import chromadb
from chromadb.config import Settings

# Initialize vector embeddings
chroma_client = chromadb.PersistentClient(
    path="./chroma_db",
    settings=Settings(anonymized_telemetry=False)
)
collection = chroma_client.get_or_create_collection(name="basketball")

# Fallback unstructured text chunks
doc1 = "LeBron injured for remainder of NBA 2023 season..."
collection.upsert(documents=[doc1], ids=["doc1"])

Using spaCy for Named Entity Recognition to drive parallel queries across the graph and vector tiers.

import spacy

nlp = spacy.load("en_core_web_sm")

def extract_entities(text):
    """ Extract entities from the given text using spaCy. """
    doc = nlp(text)
    return list(set([ent.text for ent in doc.ents]))

def get_facts(qs, entities):
    """ Retrieve facts for a list of entities from the QuadStore. """
    facts = []
    for entity in entities:
        subject_facts = qs.query(subject=entity)
        object_facts = qs.query(object=entity)
        facts.extend(subject_facts + object_facts)
    return list(set(tuple(fact) for fact in facts))

Practical Applications

• Dynamic Sports Analytics: Using Priority 1 graphs for current roster status and Priority 2 for historical stats to prevent the LM from misattributing a player’s current team. Pitfall: Relying on vector-only RAG often leads to ‘hallucinated transfers’ when multiple team names appear near a player’s name in training data.
• Automated Fact-Checking: Deploying a 3-tiered system where verified immutable truths (Priority 1) automatically override conflicting background information (Priority 2). Pitfall: Failing to provide an explicit hierarchy in the system prompt allows the LM to choose ‘statistically likely’ but incorrect answers from latent weights.

References:

• https://machinelearningmastery.com/beyond-vector-search-building-a-deterministic-3-tiered-graph-rag-system/