Jylus · Real-time data platform for applications and AI

Explore the data platform behind your application.

Ingest and retain records, query and search them, inspect state and relationships, then prepare AI evidence. Build on the same data through the API, with or without an AI integration.

Start here to see how it works. No account is needed for the synthetic demonstration or the single-use upload trial.

Run an example

Follow one order through records, search, changes, relationships and AI context.

Run an exampleNo account · real backend operations

Test my data

Paste a small sample or upload a file. Inspect readable facts and source references.

Test my dataSingle-use sample · up to 32 KB · no account

Build with the API

Create a free workspace, retain records and run a workload quickstart.

Build with the APIFree workspace · 1 GB included · no card

An application workload, end to end

Run the data behind an order-support application.

An order is created. Its shipment is reported delayed. A later correction changes that status. Follow the same records through your application's data work and into AI context.

Records

Four observations are stored and read back: an order is created, a delay is reported, a correction arrives and a customer opens a ticket.

Run the demonstration to see real responses here. No result is precomputed or substituted.

Separate recorded comparison · no model call

Compare model input and cost.

The walkthrough estimates evidence size. This separate comparison shows the published same-model benchmark: complete model input, cost per query and scored answer quality. It does not use your sample or rerun a model.

Frozen benchmark: 528 questions across four data domains
Same Gemini 3.1 Flash Lite model and settings on both paths. Average per query; different data from the four-record walkthrough.

Full context

Average complete model input: 205,129 tokens
Published model cost per query: US$0.053305
Strict answer accuracy: 78.79%
Scored hallucination trials: 7.58%

Jylus Context Pack

Average complete model input: 2,532 tokens
Published model cost per query: US$0.001160
Strict answer accuracy: 100.00% observed
Scored hallucination trials: 0.00%

The published 98.77% input reduction and 97.82% model-cost reduction apply to this defined benchmark. They are not savings measured in your trial. A small Context Pack can be larger than its source records.

What was measured, and how were answers checked?

Model input is the complete input used by the model, including the question, instructions and supplied evidence, rather than the walkthrough's source-size estimate. The benchmark compares full context and Jylus evidence using the same model settings and deterministic scorer.

The strict answer gate checks the expected answer and required source evidence. The published contract retains raw answers, provider request IDs, timings, hashes and checkpoints. Acceptance requires zero missing pairs, duplicate cells, execution errors or malformed response hashes.

Cost is the published per-query model cost calculated from provider-reported usage at the benchmark's frozen standard rates. This is not an invoice or a current quote. Your bill also depends on output, caching, rates and usage; Jylus subscriptions and other application costs are separate.

528/528 Jylus answers passed in this cohort. That is an observed result, not a universal accuracy guarantee. The evaluation has not been independently reproduced.

Read the benchmark methodology Test your own data

Single-use trial

Drop in data. Get a proof-bound result.

No account is required. Jylus creates an isolated one-time tenant, compiles the result, blocks further access immediately, and sends the tenant through the deletion lifecycle.

Start with a synthetic example:

RESULTYour compiled evidence will appear here.

The response contains source-backed facts, proof IDs, conflicts, missing evidence and the deletion state.

Start with a specific application path

What can Jylus replace?

Choose a workload to evaluate. See the current architecture, the Jylus path, what can disappear and what still belongs in your stack.

Chroma Cloud

Assess and copy a Chroma Cloud workload

Use Jylus Migrate to discover collections, assess compatibility and storage, then approve a verified record-copy evaluation while Chroma Cloud keeps serving.

Explore this replacement evaluation →

MongoDB

Evaluate an order-event read workload

Move the records and reads behind a support screen: order observations, a bounded timeline and exact field queries.

Explore this replacement evaluation →

Elasticsearch

Evaluate support-event search and investigation

Search shipment reports, filter by order and inspect the observations and linked records behind a support incident.

Explore this replacement evaluation →

Pinecone

Evaluate retrieval plus order-state evidence

Use one order-support workload to test retrieval, source lookup, corrected state and the context an AI agent receives.

Explore this replacement evaluation →

Databricks

Serve live order history to your application

Test the order-history API used by your support application, alongside your existing analytics system.

Explore this replacement evaluation →

Prefer to keep your stack? Jylus can also work alongside it as an evidence API for your existing application and model. Follow the additive integration guide →

Measured on published test workloads

These benchmark scores are separate from the data you submit in the trial.

View methodology →

Measured native stack50.695% QASPER NDCG@1075.130% QASPER Recall@1036.443% FinQA execution40.048% live TEMPO NDCG@10

Free storage1 GB

Ingest capacity1,000 events/sec

Data regionSydney, Australia

AccessHTTPS API

Continue with the API.

The free workspace includes 1 GB. Choose retention, create a scoped key and run your selected workload. Model-provider charges are separate.

Open your workload quickstart

API and AI integration details

Test AI evidence

See what your model gets when the evidence improves.

Use your own current state, history and documents. Jylus returns a bounded Context Pack with facts, timeline evidence, contradictions, freshness and proof IDs.

01Send dataUse real records, including irrelevant history and conflicting values.
02Ask Jylus for evidenceCall /api/v1/analyze with the question your model needs answered.
03Inspect the Context PackCheck retained facts, proof IDs, token estimate, completeness and execution time.
04Compare against your modelRun the same question with your current context path and the Jylus evidence.

Generate your first Context Pack →

QASPER · 1,335 TEST QUERIES50.695% NDCG@10

75.130% Recall@10 · 45.663% MRR@10

FINQA · 1,147 TEST CASES36.443% execution accuracy

31.473% program accuracy · zero model calls

TEMPO · FULL LIVE API COHORT40.048% NDCG@10Measured on the native Jylus stack.QASPER uses the untouched mteb/QASPER test task. FinQA uses the official program executor. TEMPO is the full 1,730-question public-v1 API result. Metrics measure different tasks and are reported separately.

The complete test loop

Ask for evidence. Inspect exactly what your model receives.

The response keeps facts, temporal evidence, contradictions, missing evidence and proof IDs explicit. Efficiency values are measured from your own request.

CONTEXT PACK REQUESTPOST /api/v1/analyze

curl https://api.jylus.ai/api/v1/analyze \
  --request POST \
  --header "Authorization: Bearer $JYLUS_READ_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "source": "current",
    "text": "What is the latest latency reported by edge-17?",
    "vector": {
      "text": "latest edge device response time"
    },
    "context": {
      "mode": "decision",
      "token_budget": 2000,
      "max_facts": 18,
      "max_timeline": 12,
      "include_documents": false
    },
    "scan_limit": 1000,
    "limit": 24,
    "include_results": false
  }'

ABRIDGED EXAMPLE RESPONSEapplication/json

{
  "success": true,
  "complete": true,
  "execution_mode": "native_current",
  "context": {
    "decision_context_id": "ctx_...",
    "decision_readiness": {
      "status": "ready",
      "can_answer": true
    },
    "facts": [
      {
        "statement": "edge-17 reported 12.4 ms latency",
        "proof_ids": ["evt_001"]
      }
    ],
    "timeline": [
      { "observed_at": "2026-08-21T08:00:00Z", "proof_ids": ["evt_001"] }
    ],
    "contradictions": [],
    "missing_evidence": [],
    "context_efficiency": {
      "source_tokens_estimated": 1842,
      "retained_evidence_tokens_estimated": 436,
      "estimated_token_reduction_percent": 76.33
    },
    "proof": { "event_ids": ["evt_..."], "complete": true }
  }
}

For applications that also need AI evidence

Beyond conventional RAG

Relevant content is a starting point. Inspect current and historical state, transitions, relationships, contradictions, missing evidence and source proof IDs before your model reasons.

Keep your existing data stack and model if that suits your workload. Use Jylus to prepare evidence alongside them, or evaluate the wider data platform first.

Explore the eight evidence capabilities

Semantic retrieval

Retrieve relevant records for a question and inspect the returned sources. Measure relevance against your own expected results.

Structured filtering

Scope queries to an order, customer or other supported fields. Check that unrelated records stay outside the result.

As-of state resolution

Compare the latest observation with an earlier source observation. The order-support demonstration shows the delay before its correction. Knowledge-time evaluations use the server-recorded arrival time.

State transitions

Inspect the timeline and the source records behind a status change. Keep event time and observation time explicit when evaluating late updates.

Contradiction detection

Inspect conflicts reported in the Context Pack rather than asking a model to silently choose a value. Verify the reported sources and scope.

Missing-evidence reporting

Check completeness, missing evidence and decision readiness. A successful HTTP response alone does not mean there is enough evidence to answer.

Proof IDs per fact

Follow retained facts back to their source proof IDs. The AI quickstart checks the proof set against the records it submitted.

Model-independent evidence

Inspect the returned JSON before passing it to your chosen model. Jylus prepares the evidence; model-provider calls and charges remain separate.

These are evaluation points for Jylus, not a blanket comparison with every search or RAG implementation. Run the AI evidence quickstart →

Test the data engine

From account to a real event in minutes.

01Create a free workspaceEmail, GitHub or Google. No payment step.
02Copy the one-time API keyThe key is scoped and shown once.
03Send your real payloadThe Developer hub includes HTTPS, NATS JetStream, Docker and OpenTelemetry setup.

Create free workspace →

FIRST EVENTcurl

curl https://api.jylus.ai/api/v1/events \
  -H "Authorization: Bearer $JYLUS_API_KEY" \
  -H "Idempotency-Key: try_evt_001" \
  -H "Content-Type: application/json" \
  -d '{
    "stream": "my-workload",
    "events": [{
      "id": "evt_001",
      "type": "telemetry.sample",
      "occurred_at": "2026-08-21T08:00:00Z",
      "data": { "device": "edge-17", "latency_ms": 12.4 }
    }]
  }'