EXPLORE OPERATIONAL DATA

InvestiGenie

Explore your data. Investigate a question.

Explore across seven data connectors. Investigate Elasticsearch or ClickHouse Data Views with up to eight sampled records, source-side aggregations, and citations you can inspect.

Explore the code How it works

Open source · Bring your own data · Choose your AI provider

CONNECTEXPLOREINVESTIGATE

A LOOK INSIDE THE APP

From question to inspectable evidence.

APP PREVIEW

Synthetic InvestiGenie walkthrough: open Explore, search checkout logs and inspect the graph, investigate one record, then review the Genie hypothesis. — Synthetic workflow · Explore → graph → investigate a record → Genie result

INVESTIGATION RESULT

Keep samples, counts, and hypotheses distinct.

The example shows bounded sample coverage alongside cited records and a broader aggregation count. Findings remain hypotheses for a person to verify.

Read about the evidence pipeline

THE IDEA

Explore the full query result. Investigate with a bounded sample.

Explore pages through matching records. AI investigations instead examine up to eight sample rows per selected Data View, using at most three bounded queries for latest records, earliest records when useful, and an unfiltered fallback if proposed filters return no rows. Source-side aggregations can count a wider result set. Results show sampled rows, aggregation counts, and cited records separately.

A CLEAR PATH THROUGH THE DATA

Connect. Explore. Investigate.

01 — 03

Connect your data

Explore Elasticsearch, ClickHouse, BigQuery, Loki, Prometheus, VictoriaLogs, and VictoriaMetrics. Credentials stay on the server and are encrypted.

SOURCES

Explore matching records

Use saved Data Views, time windows, source-side sorting, histograms, and paging to inspect query results.

DISCOVER

Investigate a question

Choose Elasticsearch or ClickHouse Data Views and a saved AI profile. The server validates structured filters, examines at most eight sample rows per Data View in at most three queries, and can request broader source aggregations. Other connectors support Explore but not AI query translation yet.

EVIDENCE

BUILT AROUND INSPECTION

Useful answers should come with a trail.

Samples are not a full scan

AI sees at most eight sampled record rows per Data View. The latest/earliest sample and empty-filter fallback are bounded; this is not statistically representative.

Counts have different scopes

Source aggregations may count more records than were sampled. A finding cites distinct, validated sample records—not the aggregation total.

Hypotheses need checking

Citations make examples traceable; they do not prove an AI conclusion. Inspect the query scope and records before acting.

UNDER THE HOOD

A small map of the system.

DOCS

ARCHITECTURE

One API connects the UI, data sources, storage, and AI.

The Rust server serves the browser app, manages encrypted saved objects, and coordinates read-only connectors and the configured AI provider.

flowchart LR
  UI["Browser UI"] --> API["Rust API"]
  API --> STORE[("Encrypted storage")]
  API --> CONNECTORS["Read-only connectors"]
  CONNECTORS --> SOURCES[("Data sources")]
  API --> AI["Configured AI provider"]

Full architecture docs

EVIDENCE FLOW

Bounded samples support hypotheses; they do not prove them.

Per Data View, the backend runs up to three bounded queries, retains at most eight unique sample rows, and keeps aggregation totals separate from citations.

flowchart TD
  Q["Question and Data Views"] --> P["Metadata and validated plan"]
  P --> L["Latest sample: up to 4 rows"]
  L -->|More matches indicated| E["Earlier sample: up to 4 rows"]
  L -->|Filters return no rows| F["Fallback in same time range"]
  L --> R["Up to 8 unique retained rows"]
  E --> R
  F --> R
  L --> A["Source aggregations"]
  R --> H["Hypotheses with cited sample IDs"]
  A --> H

Full investigation docs

BUILDS AND DOWNLOADS

Latest stable release

Release page

Loading stable release…

A NOTE ON TRUST

Evidence is still data.

The question, field metadata, compact sample records, and aggregation summaries are sent to the configured AI provider. Connection credentials and source URLs are not. Review provider retention terms before sending sensitive data. Record text is untrusted input; citations make examples traceable, not conclusions correct. Results are hypotheses, not verified diagnoses. The server has no built-in authentication or authorization; put it behind a trusted authenticated gateway.

INVESTIGATE WITH CONTEXT

Start with the data you have.

Visit GitHub