Hybrid Query Benchmarks¶

This benchmark focuses on the query shape that matters most for NornicDB's positioning: semantic retrieval followed by graph expansion in the same engine.

The goal is not to claim a universal leaderboard result. The goal is to show what happens when vector search and one-hop graph traversal share one execution path instead of being stitched together across multiple systems.

Summary¶

• Vector-only queries stayed in sub-millisecond to low-millisecond territory locally, depending on transport.
• Vector + one-hop graph traversal added a small incremental cost locally.
• Full RRF retrieval averaged 19.47 ms over Bolt and 20.40 ms over HTTP on the production-scale fixture.
• Full RRF retrieval + one-hop graph traversal averaged 20.00 ms over Bolt and 20.28 ms over HTTP.
• Remote latency tracked client-to-server RTT, which means end-to-end latency became network-bound rather than database-bound.

Original 67K Direct-Vector Benchmark¶

Test Setup¶

Local environment:

• Apple M3 Max
• 64 GB RAM
• Native macOS installer

Remote environment:

• GCP
• 8 vCPU
• 32 GB RAM

Local Results¶

Traversal Queries¶

Bolt is nearly zero allocation. This was measured under concurrent load with mixed HTTP and Bolt queries. The tail-latency spikes came from garbage collection triggered by the HTTP path. Bolt is more efficient than HTTP for tail latency.

Remote Results¶

Client-to-server latency was about 110 ms.

Once local direct-vector compute is in the sub-millisecond range, network RTT dominates the user-visible latency budget.

Larger 300K-Node Full-RRF Benchmark¶

Run on 2026-08-28 at revision c27fff0fbfd2 using the real BGE-M3 fixture in data/test-200kembed. The writable data directory was cloned from that source before mutation. The benchmark waited for the search readiness endpoint to report ready and verified the node and edge counts before measuring.

Test Setup¶

Local environment:

• Apple M3 Max
• 64 GB RAM
• Source build with the localllm tag

The server used no search-result or embedding cache and had zero background embedding workers. Every request used a unique query cache key, so every sample executed vector retrieval, BM25 retrieval, and RRF fusion. Each response asserted search_method = rrf_hybrid, a positive RRF score, at least one source rank, and fallback_triggered = false; the hop shape also required a neighbor.

Results¶

Reproduce the focused run:

NORNICDB_LARGE_RRF_E2E=1 \
NORNICDB_LARGE_RRF_HTTP_ADDR=127.0.0.1:17474 \
NORNICDB_LARGE_RRF_BOLT_ADDR=127.0.0.1:17687 \
NORNICDB_LARGE_RRF_DATABASE=translations \
NORNICDB_LARGE_RRF_EDGES=1000000 \
NORNICDB_LARGE_RRF_WARMUP=3 \
NORNICDB_LARGE_RRF_ITERS=800 \
go test -tags=e2e ./testing/e2e \
  -run '^TestLargeDatasetRRFRoundTrip_BoltVsHTTP$' -count=1 -v

The address variables attach the test to an already configured benchmark server. Omit them to let the test copy-on-write clone data/test-200kembed, build the local-embedding binary, and manage the server itself.

Why This Matters¶

Most systems make this query shape a composition problem:

1. embed the query
2. call a vector store
3. move the results into a graph store or application layer
4. expand neighbors and shape the result there

NornicDB keeps that inside one execution engine. The benchmark does not prove every workload is constant-time, but it does show that shallow hybrid retrieval can stay tight enough locally that deployment topology matters more than extra database-side micro-optimizations.

Caveats¶

• These are single-node measurements.
• The dataset is not billion-scale.
• Remote throughput is latency-bound, not compute-bound.
• These numbers are useful for query-shape comparison, not as a blanket claim for every graph or vector workload.

Verification Queries¶

Vector-only:

curl -s -u "$NORNIC_USERNAME:$NORNIC_PASSWORD" "$ENDPOINT" \
  -H "Content-Type: application/json" -H "Accept: application/json" \
  -d '{
    "statements":[
      {
        "statement":"CALL db.index.vector.queryNodes('\''idx_original_text'\'', $topK, $text) YIELD node, score RETURN node.originalText AS originalText, score ORDER BY score DESC LIMIT $topK",
        "parameters":{"text":"get it delivered","topK":5},
        "resultDataContents":["row"]
      }
    ]
  }'

Vector + one-hop graph traversal:

curl -s -u "$NORNIC_USERNAME:$NORNIC_PASSWORD" "$ENDPOINT" \
  -H "Content-Type: application/json" -H "Accept: application/json" \
  -d '{
    "statements":[
      {
        "statement":"CALL db.index.vector.queryNodes('\''idx_original_text'\'', $topK, $text) YIELD node, score MATCH (node:OriginalText)-[:TRANSLATES_TO]->(t:TranslatedText) WHERE t.language = $targetLang RETURN node.originalText AS originalText, score, t.language AS language, coalesce(t.auditedText, t.translatedText) AS translatedText ORDER BY score DESC, language LIMIT $topK",
        "parameters":{"text":"get it delivered","topK":5,"targetLang":"es"},
        "resultDataContents":["row"]
      }
    ]
  }'

Full-RRF Verification Queries¶

Full RRF retrieval:

curl -s -u "$NORNIC_USERNAME:$NORNIC_PASSWORD" "$ENDPOINT" \
  -H "Content-Type: application/json" -H "Accept: application/json" \
  -d '{
    "statements":[
      {
        "statement":"CALL db.retrieve($request) YIELD node, score, rrf_score, vector_rank, bm25_rank, search_method, fallback_triggered RETURN node.originalText AS originalText, score, rrf_score, vector_rank, bm25_rank, search_method, fallback_triggered ORDER BY rrf_score DESC LIMIT 5",
        "parameters":{"request":{"query":"chain baseline","embedding":[0.95,0.05,0.0],"limit":5,"types":["OriginalText"]}},
        "resultDataContents":["row"]
      }
    ]
  }'

Full RRF retrieval + one-hop graph traversal:

curl -s -u "$NORNIC_USERNAME:$NORNIC_PASSWORD" "$ENDPOINT" \
  -H "Content-Type: application/json" -H "Accept: application/json" \
  -d '{
    "statements":[
      {
        "statement":"CALL db.retrieve($request) YIELD node, score, rrf_score, vector_rank, bm25_rank, search_method, fallback_triggered MATCH (node)-[:BENCH_HOP]->(neighbor:BenchmarkHop) RETURN node.originalText AS originalText, elementId(neighbor) AS neighborID, score, rrf_score, vector_rank, bm25_rank, search_method, fallback_triggered ORDER BY rrf_score DESC LIMIT 5",
        "parameters":{"request":{"query":"chain baseline","embedding":[0.95,0.05,0.0],"limit":5,"types":["OriginalText"]}},
        "resultDataContents":["row"]
      }
    ]
  }'

Related Reading¶

• Current Northwind Benchmark vs Neo4j
• Retrieval Recall Benchmark
• Graph-RAG: Typical Distributed vs NornicDB In-Memory
• Canonical Graph + Mutation Log Guide