Core module · APIXX.ai

APIXX AI

APIXX AI is the operational intelligence engine for your integration estate. It master-queries APIXX.flow (pipelines, runs, errors) and APIXX.data (canonical records, coverage, quality) and returns structured, evidence-backed analyses — never guesses.

What APIXX AI is

Most AI assistants generate plausible text. APIXX AI generates operational intelligence: a single structured report that names the specific flows, errors, and records involved in whatever you asked about, with a confidence score that reflects how strongly your workspace data supports the conclusion.

Workspace-grounded, not web-grounded

Every answer is built from your own flows, runs, errors, connectors, and health telemetry. APIXX AI never invents flow names, error codes, or counts.

The three modules

The APIXX platform is organized into three peer modules, each with its own sidebar:

Module	Path	Purpose
APIXX.flow	/flows, /runs, /errors	Design, schedule, and operate integration pipelines.
APIXX.data	/data/dashboard	Canonical data layer — entities, coverage, quality, provenance.
APIXX.ai	/apixx-ai	Operational intelligence over both modules.

Access & workspace

Open APIXX AI from the module switcher (top-left) or go directly to /apixx-ai. Any authenticated user in your workspace can ask questions; the AI only ever sees the customer record bound to your session — there is no cross-tenant leakage.

Asking questions

The composer accepts up to 2,000 characters. Press Enter to send, Shift+Enter for a newline. Each submission produces one analysis card on the right and a history entry on the left, so you can pivot between past answers without re-running them.

Examples that work well:

"Why are orders delayed today?"
"Which flows are failing most often in the last 24 hours, and what's the dominant error?"
"Are any connectors at risk of disconnecting?"
"Summarize integration health right now."
"Why does the NetSuite customer push keep partially failing?"

Anatomy of an analysis

Every analysis card returns the same six sections:

Section	What it contains	How to read it
Summary	1–3 sentence plain-language answer.	Skim first; the rest is evidence.
Confidence pill	0–100 score.	See the Confidence scoring section below.
Root Cause	Explanation plus factor chips.	The chips are the contributing signals (rate-limited connector, mapping change, etc.).
Supporting Data	Labeled metrics with rationale.	Each row cites a number pulled from your workspace and why it matters.
Event Timeline	Ordered events with when / title / detail.	Reconstructs the sequence that led to the current state.
Recommendations	Concrete next steps inside the app.	Each step links to a place you can act — pause a flow, retry a run, re-authorize a connector.

Confidence scoring

Score	Meaning	What to do
90–100	Clear single cause backed by multiple data points.	Act on the recommendations.
70–89	Likely cause with some ambiguity.	Verify against the linked runs before acting.
50–69	Plausible hypothesis, weak evidence.	Treat as a starting point; gather more telemetry.
Below 50	Insufficient data — the summary will say so explicitly.	Re-run after the next scheduled flow execution.

Watchtower

The right pane runs a continuous lightweight scan over three signals — flow status, 24h run failure rate, and connector connectivity — and rolls them up to nominal, warning, or critical. Click Scan to force a refresh. Watchtower is deterministic (no AI), so it's safe to wire alerts off it.

Data the AI can see

For each question, APIXX AI loads a bounded context window from your workspace:

All flows (status, names, error counts).
The last 24 hours of runs, capped at 500 entries.
All connector profiles and their current status.
The most recent system health snapshot.
An aggregated top-error table (flow × message) derived from the runs above.

No record-level PII

The AI never receives raw canonical records, customer email addresses, or payment data — only metadata about flow execution. Direct record-level questions are answered with guidance to query APIXX.data instead.

Limits & guarantees

One question = one analysis. Multi-turn chat is intentionally not supported; each query is self-contained and reproducible.
Structured JSON contract. Responses are schema-validated; malformed model output is rejected with a clear retry prompt rather than rendered.
Bounded context. Run history beyond 500 records or 24 hours is summarized, not enumerated.
No writes. APIXX AI never pauses flows, retries runs, or modifies data — it only recommends. You stay in control.
Auditability. Every analysis carries a generatedAt timestamp.

Tips for good answers

Name the symptom, not the cause. "Orders are delayed" beats "Shopify webhook is broken" — let the AI find the cause.
Anchor in time. "…in the last 4 hours" or "…since yesterday" sharpens the analysis.
Reference flow or connector names if you already suspect one, but don't prescribe the answer.
Re-ask after acting. Run a follow-up question to confirm the recommendation worked.
Low confidence is a signal. Treat sub-60 scores as "more telemetry needed," not as a hedge.

PreviousAPIXX Data Next Usage & quotas