Analytics system flips AI chat from question-first to analyst-first
The paper starts from a familiar failure: conversational analytics tools assume the user already has a well-formed question, which leaves a non-expert staring at a blank query box on a schema they do not understand. Existing fixes only narrow the gap. Commercial 'proactive' tools detect statistical anomalies, but only over metric layers that an analyst has already curated by hand. Academic next-question recommenders need query logs to learn from, and a fresh dataset has none. The authors describe a production system that inverts the model from question-first to analyst-first, built on two coupled ideas. The first is a pluggable domain-expert 'skill': a self-contained, folder-based, database-free subject-matter pack holding a manifest, per-stage prompt facets, keyword-routed references, report templates and optional compute. A skill is auto-selected for each client-and-dataset pair by deterministic schema matching, then spliced as a cross-cutting concern into every stage of the agentic pipeline, the schema explorer and the report engines; when no matching skill exists, it degrades to a strict no-op rather than failing. Because each skill is just a folder resolved deterministically, the catalogue can grow without limit, forming what the authors call an extensible marketplace of domain experts. The second idea is an offline knowledge-compilation loop. An agent probes the dataset's parquet files through DuckDB, adding zero load to production, and runs critic-gated convergence table by table with self-healing retries, validating joins by checking value overlap between columns. The result is durable schema knowledge that drives standing expert reports, where every published metric is re-verified by re-executing the SQL that produced it, plus a set of suggested questions that mirror the report's own agenda. Together these close a proactive loop: a report surfaces numbers, the numbers seed follow-up questions, and clicking one launches an already-verified deep dive, all before anyone touches the query box. The paper presents a formal model of the architecture and illustrative evidence from a single tenant, but states explicitly that it makes no user-study or benchmark claims; the authors frame the contribution as the architecture itself and its defensibility, not a measured performance result.
Key facts
- The architecture rests on two coupled pieces: a pluggable domain-expert 'skill' abstraction and an offline knowledge-compilation loop.
- Skills are self-contained, database-free folders (manifest, prompt facets, references, report templates, optional compute) auto-selected per client-and-dataset by deterministic schema matching, and they degrade to a strict no-op when no match exists.
- The compilation loop has an agent probe the dataset's parquet via DuckDB at zero production load, run critic-gated per-table convergence with self-healing retries, and validate joins by value overlap.
- Every metric in the resulting standing reports is re-verified before publication by re-executing the evidence SQL that produced it, and the reports also generate suggested follow-up questions.
- The paper explicitly disclaims user-study or benchmark results, offering only a formal model and illustrative single-tenant evidence as support for the architecture.
Why it matters
Conversational analytics has a cold-start problem: a non-expert user cannot ask a good question about a schema they have never seen, and the two existing workarounds do not actually solve it. Anomaly-detecting 'proactive' tools still need an analyst to have built the metric layer first, and next-question recommenders need query-log history a new dataset does not have. This architecture tries to remove the blank query box entirely by generating the useful reports and questions before any user interaction, from schema knowledge the system compiles for itself.
Who it affects
The design targets teams building or buying enterprise analytics products for clients with unfamiliar, fresh datasets and no history of prior queries to learn from. Because the skill packs are self-contained folders resolved by schema matching rather than hardcoded per client, the same core pipeline is meant to serve many different subject domains through an expandable catalogue of skills, described as an open-ended marketplace.
How to use it
A domain-expert skill is dropped in as a folder and picked up automatically for any client-dataset pair whose schema matches it, requiring no database of its own; if none matches, the system falls back cleanly rather than breaking. The knowledge-compilation loop runs offline against the dataset's parquet files via DuckDB before any user-facing interaction, producing the durable schema knowledge that standing reports and their suggested questions are built from at query time.
How solid is it
The paper is upfront about its evidentiary limits: it offers a formal model of the architecture plus illustrative evidence from a single tenant, and states directly that it makes no user-study or benchmark claims. The claimed contribution is the architecture and its defensibility as a design, not a measured comparison against alternatives.
Risks and caveats
No authors, affiliations or institution are identified for the system or the paper, no client, industry or dataset domain is named for the production deployment described, and no underlying model or LLM powering the agentic pipeline is specified. With evidence drawn from a single tenant and explicitly no benchmark or user-study data, how the approach performs at scale or across domains remains unshown by the paper itself.
“We give a formal model and report illustrative single-tenant evidence. We make no user-study or benchmark claims; the contribution is the architecture and its defensibility.”
— the paper