DATA & AI · BY GUSTAVO

Project
Summary

Explore the business question and the implementation behind this project.

Open assistant

Business Answers That Can Be Checked Against Their Sources

What I Built and How I Checked It

Regression Results, SQL Evidence and Response Controls

01 / FIXED REGRESSION BENCHMARK14 of 14 checks passed

SQL, Retrieval and Request Boundaries All Pass Their Fixed Checks

Each tile represents one predefined regression case. The recorded evaluation checks exact result values, expected source documents and appropriate clarification or refusal behavior.

Structured analytics5 / 5 passed

Sales total, peak month, department count, total workforce and a monthly sales filter.

Document retrieval5 / 5 passed

PTO carryover, remote-work policy, campaign audience, API rate limits and a Spanish-language policy question.

Request boundaries4 / 4 passed

Unavailable periods, unsupported breakdowns, unrelated questions and a request for a secret.

This is a 14-case regression benchmark. Gemini is disabled during evaluation. The result verifies these fixed behaviors; live answer quality and latency require separate measurement.

Recorded CI evaluation · Cases and assertions. The same run passed PostgreSQL integration and Docker startup checks.
02 / THE ANSWER JOURNEYExplore each step

Hybrid Search and SQL Support Different Questions

Business metrics use parameterized, read-only SQL queries. Policy and product questions combine semantic and full-text search to retrieve supporting passages.

Architecture from the repository. Gemini is optional; Python routes questions deterministically.
03 / BUSINESS DATA IN ACTION2024 demo records

The Sales Answer Is Computed from 60 Versioned Records

The sales table contains 60 product/month records. For Crestline X, SQL returns 112,182 units across 2024 and identifies November as the peak at 12,974. Select a month below to inspect its value. This is an example of a verifiable answer, not a business-impact metric.

November · 12,974 units · highest month

Source: Versioned business dataset. Fictional company records; these are example answers, not real-company performance.
04 / KNOW WHAT IS ANSWERINGTransparent response modes

Invalid or Unavailable Generation Falls Back to Evidence

The user sees which response mode produced the answer. Gemini can summarize retrieved passages or computed rows; citation and numeric checks reject specific unsupported output. If generation is unavailable or invalid, the assistant returns concise cited evidence.

This explains application behavior; it does not report the live demo’s current API availability. No accuracy percentage is implied.

A Small Corpus with Inspectable Query and Retrieval Paths

01

Structured business records

60 sales records, 20 campaign records, 12 monthly metrics and six department aggregates representing 100 employees. Dataset fingerprints detect changed fixtures.

02

Hybrid document retrieval

Nine source documents are searched using 384-dimensional multilingual embeddings, exact cosine search and PostgreSQL full-text ranking. English and Spanish questions are supported.

03

Controlled execution

SQL uses bound parameters, read-only templates and a five-second statement timeout. Explicit follow-ups retain relevant session context. Users can inspect the query and supporting evidence.

Architecture and retrieval details

A Working Demo with an Explicit Evaluation Boundary

    Explore on GitHub

    Reviewed 22 October 2026 against the project README and linked evaluation reports. Figures describe the documented dataset or test run.

    BEHIND THE EXECUTIVE SUMMARY

    Technical details