Jylus

Jylus is an evidence and state layer for AI systems that need to reason over data which keeps changing. It sits between operational records and the model, retrieving the relevant current state, history, relationships, and semantic matches before compiling them into a compact Context Pack. Each retained fact carries proof IDs tied to source records, while contradictions, freshness, and missing evidence stay explicit. The point isn't to generate the final answer. Jylus prepares bounded, traceable input that a team's chosen model can use, so the model spends its effort reasoning over evidence instead of searching a large and possibly stale body of data.

The product addresses a gap that ordinary vector retrieval doesn't handle especially well. Similarity search can find text that resembles a question, but it may not know which record is current, whether a late correction changes an earlier event, or what was known at a particular time. Jylus resolves current and historical state, tracks transitions, follows relationships, and reports conflicts rather than flattening them away. Its analysis endpoint accepts a question plus controls such as token budget, fact and timeline limits, and an optional distinction between when something happened and when the evidence became known. The returned Context Pack includes facts, timelines, proof references, completeness, missing evidence, and estimates of how much source material was reduced.

Teams feed Jylus through an authenticated HTTPS API, with managed NATS JetStream available on paid plans for higher throughput. Events support idempotency keys, which lets an application retry ingestion without duplicating accepted records. Scoped API keys, expiry, rate limits, audit events, and workspace query limits give operators controls around the data path. The developer surface also includes API documentation, quickstarts for people and coding agents, and OpenTelemetry setup. A control plane shows streams, API keys, billing, support, ingest rate, response time, retained storage, and system health. Jylus can therefore serve as both the retrieval layer used by an application and an operational surface for watching that layer.

It fits developers building agents, model-backed applications, security investigations, observability workflows, and operational decision systems where the answer depends on changing state. A support agent could need the latest account status plus the timeline that led there. An incident workflow could combine a current service condition with recent telemetry, related infrastructure, and a runbook. A prediction system could compare a live signal with relevant historical cases without sending every stored event to the model. Jylus keeps the evidence stage model independent, so teams can retain their existing model provider and data stack while adding one API before the reasoning step.

What separates Jylus from a generic RAG pipeline is the emphasis on temporal state and proof. The service exposes as-of state resolution, state transitions, structured filtering, relationship traversal, contradiction detection, and missing-evidence reporting alongside semantic retrieval. It also publishes test methodology and measured results for defined workloads, including a frozen operational benchmark and separate QASPER, FinQA, and TEMPO evaluations. Those measurements should be read in the context of their stated workloads rather than treated as a universal accuracy claim, but publishing the method makes the product easier to assess. A no-account trial lets a visitor submit up to 32 KB of text, JSON, or log data to inspect a one-time Context Pack before creating a workspace.

The security posture is designed around an infrastructure product. The site describes server-side sessions, one-time API key reveal, least-privilege scopes, Argon2id password hashing, HttpOnly cookies, HMAC-hashed API secrets, CSRF protection, restrictive browser policies, and idempotent event writes. The trial creates an isolated one-time tenant, blocks further access after compilation, and sends it through the deletion lifecycle. Jylus also says submitted data isn't used to train general-purpose AI. These are useful controls, though teams with formal procurement requirements will still need to review the service terms, regional hosting, subprocessors, retention limits, and any contract-specific deployment or support commitments.

Jylus uses freemium pricing. The free Developer plan includes 1 GB of searchable storage, seven days of searchable history, four concurrent queries, HTTPS access, and throughput up to 1,000 events per second per workspace without a credit card. Paid self-service tiers raise throughput, storage, history, concurrency, and transport options, starting with Builder at A$59 per month, followed by Startup at A$299 and Scale at A$1,999. Enterprise agreements cover custom capacity, architecture review, named support, and contract-defined terms. It is a focused infrastructure choice rather than a ready-made chatbot, and it makes the most sense when traceable evidence from live data is important enough to justify operating a dedicated context layer.