Skip to content

Tuning Reference

treedinteractive edited this page Jul 6, 2026 · 2 revisions

Tuning reference

SERAPH runs well on its defaults. These are the knobs that matter when you want to tune for your own corpus:

Setting Default What it controls
tau auto Relevance threshold — the main speed ↔ completeness dial. Higher = tighter, faster; lower = broader, deeper.
top_k 10 Number of results returned.
eigen_threshold 10 How readily hub points form — shapes the graph backbone.
search_entry_k 8 Minimum entry points per query.
search_entry_gap 0.05 Widens entry selection automatically for ambiguous queries.
search_max_depth 50 Traversal depth limit.
search_stream_threshold 50,000 Store size at which search switches to the streaming beam.

A note on measuring quality. SERAPH is deliberately not a nearest-neighbour index, so don't grade it against brute-force cosine recall — that measures the one thing it's designed not to be. The honest test is reading the results: retrieval routes by intent, and the right answer is the one the geometry reaches, not necessarily the closest vector.


Performance at a glance

Measured on a single RTX 4070 (12 GB), bge-small 384-dim encoder, a 900,000-item Wikipedia corpus.

Operation Result
Search latency (900k items) ~0.9 ms typical query
Ingest ~1,400 items/sec, flat from 1k to 900k
Encoding (GPU) ~1,300/sec on full paragraphs, up to ~8,500/sec on short text
Integrity verify ~1.2M items/sec — a 900k store audits in under 1 second
Federation routing ~0.005 ms to route a query to the right store
Provenance depth 5–12 hops to origin (short, verifiable chains)

A few things worth calling out:

  • Search is sub-millisecond and stays nearly flat to 100k items, rising only gently beyond that. It scales because entry selection runs on the GPU and the graph walk only ever touches the relevant region — not the whole store.
  • Ingest throughput is flat across three orders of magnitude. Bulk-building a large store doesn't slow down as it grows.
  • Integrity is cheap enough to run routinely. The whole store re-verifies from its cryptographic chain in well under a second.

Clone this wiki locally