Benchmark with FAISS baselines

The DCEE repository includes benchmark_dcee.py, which measures Recall@K against exact inner-product neighbors (normalized vectors = cosine similarity) on the same queries for every method.

Example snapshot

Representative run: 50,000 normalized vectors, same 200 queries, K = 5, DCEE with tuned + AMP settings. Numbers depend on machine, drivers, and dataset — treat as directional, not guarantees.

Method	Recall@5	P50 (ms)	P95 (ms)	QPS (approx.)	Build (s)	Size (MB)
DCEE	96.4%	0.97	1.01	422	8.57	6.40
FAISS IndexFlatIP	100.0%	0.53	0.79	1897	0.01	25.60
FAISS HNSW (M=32, ef=64)	100.0%	0.09	0.11	10689	0.63	39.21
FAISS IVF-Flat (nprobe=8)	90.6%	0.03	0.03	36364	0.48	26.47

How to read this

Flat IP is the exact baseline (full stored vectors, brute-force search per query in this harness).
HNSW / IVF are mature approximate indexes — often faster query times with different memory footprints.
DCEE emphasizes smaller on-disk index size (quantized deltas) with approximate recall; tune probes and refinement for your SLA.

Compression

With int8 delta quantization, payload is often around ~4× smaller than raw float32 storage for the vector data — exact ratio depends on settings and corpus correlation.

Multi-hop retrieval benchmark

Besides single-hop Recall@K, DCEE can also be used in iterative multi-hop retrieval loops where each round expands the frontier with the top-K neighbors of all current hits (a common pattern in graph-style RAG systems).

The repository includes benchmark_multihop_retrieval_dcee.py, which simulates synthetic chains of correlated embeddings and measures how often a distant target node becomes reachable within a small number of hops under the same expansion policy for both DCEE and an exact cosine oracle.

Chain length L	Exact multi-hop%	DCEE multi-hop%	Exact hops (avg)	DCEE hops (avg)
2	100.0%	100.0%	1.00	1.02
3	100.0%	100.0%	2.00	2.12
4	100.0%	100.0%	2.23	2.24
5	100.0%	100.0%	2.26	2.32

On this synthetic multi-hop benchmark (chain length 2–5, beam = 32, max depth = 8), DCEE matches the exact cosine oracle with 100% multi-hop recall while keeping multi-hop expansion in the tens of milliseconds per batch, showing that the compressed index works reliably for multi-hop retrieval–style expansion.

TurboQuant-style GloVe findings (DCEE)

Using benchmark_turboquant_style_dcee.py on GloVe dimensions 50/100/200/300 (100k base, 1k queries), DCEE shows a strong bits-per-vector efficiency profile while maintaining practical Recall@10 and throughput.

Dim	Recall@10 (%)	Build (s)	Query (s)	QPS	Bits / vec	Est. MB
50	94.51	5.3732	3.8229	261.6	512.0	6.40
100	92.59	9.7500	5.2909	189.0	992.0	12.40
200	90.33	23.1313	6.4738	154.5	1952.0	24.40
300	87.99	41.1703	12.1304	82.4	2912.0	36.40

Bits/vec finding: this run shows a strong, predictable compression profile (bits/vec scaling linearly with dimension).
Next dataset focus: DBpedia benchmark runs are in progress to compare with published TurboQuant-style settings at higher dimensions.
Next optimization focus: improve retrieval quality (Recall@10) while preserving DCEE's compact bit budget.