HERMES LABS / SUPERSEARCH
Evidence infrastructure for agents

SEARCH WIDELY.
KEEP THE RECEIPT.

SuperSearch turns public discovery into an inspectable packet: what was searched, what was read, what failed, and which passages an agent actually receives.

60-second product replay

One objective. A visible chain of custody.

Frozen demonstration data, not a live search or benchmark. The replay exists to show the contract without exposing the private compiler or implying a hosted API is available.

“Map practical SQLite vector-search options for an embedded agent.”
01 / DISCOVER3 candidatesURLs retain every retrieval-route stamp.
02 / READ2 pagesA discovered URL is not mislabeled as a fetched page.
03 / SELECT3 passagesSmall, attributable evidence units enter the packet.
04 / REPORT1 timeoutThe failed paper source remains visible.
WEB · COMPLETEDCODE · COMPLETEDCOMMUNITY · COMPLETEDPAPERS · TIMED OUT
P01 / INSTALLATION
“sqlite-vec adds vector search without requiring a separate database service.”
PAGE READ · SOURCE: sqlite-vec repository · WEB / CODE / COMMUNITY
P02 / USAGE
“Load the extension before creating a virtual table for storing and searching vectors.”
PAGE READ · SAME PAGE, DISTINCT PASSAGE · CONTENT HASH RETAINED
P03 / PORTABILITY
“USearch presents a single-file similarity-search engine with bindings for several languages.”
PAGE READ · SOURCE: USearch repository · CODE / WEB
WHY THIS MATTERS

The packet can be partial and still useful. Missing evidence stays missing; a model never gets permission to silently fill the gap.

Free and open

Retrieval kernel

  • Local CLI and Python API
  • Deadline-bounded source fan-out
  • Source-state and provenance receipts
  • Schema, tests, and agent integration examples
Private preview

Evidence compiler

  • Objective-to-query-family planning
  • Managed fetching and passage custody
  • Coverage and gap accounting
  • Optional cited synthesis and operational controls

Use the kernel. Shape the service.