Interactions Analytic

Interactions is the Data Agents configuration type for ecosystem.Ai recommender logging. It pairs recommendation calls with attributed responses, then publishes family reports (markdown, notebook, graphs, JSON).

It is not the Runtime scorer. Scoring still uses POST /invocations then POST /response. Interactions reads those logs after the fact. It is not Financial ARPU/ROI and not Spend Personality scoring.

When to use it

Use Interactions when you have Runtime-style collections, typically:

Optional: offer-matrix snapshots, customer/channel collections. Do not point this type at a generic sales ledger (use Financial).

After a Workbench simulation or production traffic, run an Interactions job on the same logging collections to inspect funnel, rank uptake, and catalog quality.

Event concepts

• Ranked recommendations in final_result are the impression denominator. There is no separate “rendered impression” flag.
• Attributed responses are UUID-joined accepts inside the attribution window (default 168 hours) after date_log.
• Calls without a response stay in the funnel (right-censored). Orphan responses are reported, not silently dropped.
• Generated-but-not-ranked candidates are not in these logs.

Setup

1. Create New → configuration type Interactions.
2. Bind Mongo (or Trino extract-only) on Source. Database and connection come from that source.
3. On the Interactions block, set Calls collection and Responses collection.
4. Click Discover Predictors, then select one or more predictors or Analyze all predictors. A job will not start until a selection is confirmed (Mongo). Trino extracts default to all predictors present in SQL.
5. Optional: Date from / Date to, attribution window (hours), declared A/B experiment groups (control campaign + treatments, randomization unit customer or call).
6. Enable families on Analysis & LLM (see below). Marketing mix is off by default.
7. Save, then Run.

Discovery also shows a scale and index preflight: call/response counts and whether exact mode is feasible. Missing source indexes are recommended and may be created at job start if missing. The analyzer never loads the full interaction population into pandas; large logs use projection, partitions, and mergeable aggregates.

Families and reports

UI family checkboxes do not all get their own markdown file. Several land inside general or offers.

Each family usually also writes *_notebook, *_results JSON, and *_graphs. Open them from Reports (/reports) or the dashboard at /recommender.

How processing works

1. Join — exact UUID attribution from calls to responses inside the window.
2. Normalize — campaign, predictor, ranked offers, channels, customer keys (customer_id, customerId, client_id, party_id, …).
3. Aggregate — funnel, campaigns, ranks, catalog snapshots, cohorts. Customer IDs are pseudonymized; cells below minimum_cell_size (default 10) are suppressed.
4. Interpret — optional LLM on compact evidence packets. The model may explain findings and experiment proposals; it cannot change metrics, priority scores, or guardrails.
5. Publish — additive recommender/v1 JSON plus markdown/notebooks/graphs. Large detail stays in paginated exports / family JSON, not the consolidated dashboard blob.

Privacy and fidelity

• Funnel totals (calls / responded / accepted) are exact over the scanned population.
• Customer profiles may be sampled or truncated; the report’s Data fidelity block says so.
• Distinct-customer / cohort counts prefer Mongo $group (exact) over the tracked sample.
• LLM sees compact packets only — not raw customer identifiers.

View results

Dashboard tabs: General, Offers / Catalog, Customers, Market intelligence, Channels, Experimentation, Marketing mix, Data Quality. Channels and experimentation views read the general/offer payloads.

Related

• Setup
• Process jobs
• All report types
• Simulations — generate call/response traffic to analyze here
• Runtime closed loop