Skip to content

Project

Changelog

What changed in each release.

v0.1.0 (unreleased)

Section titled “v0.1.0 (unreleased)”

The first release: files in, a verified and cited report out.

  • Ingest and profile.
    • Spreadsheets as real tables: region detection, merged headers, totals, several tables per sheet.
    • CSV; PDF, DOCX, EML, MD and TXT sections; an optional Docling path for hard PDFs.
    • Profiling with data-quality findings (unit errors, ambiguous dates, duplicates, gaps) and entity matching.
  • Metric layer. Definitions drafted (by a model or by deterministic fallbacks), validated by code, shown in plain words, confirmed or kept as flagged assumptions. Each compiles to parameterized SQL.
  • Analysis.
    • Ten insight templates with noise bands and materiality.
    • Contribution, price–volume–mix, and rate vs mix with a Simpson guard.
    • Impact ranges computed in code.
    • Ranking with one slot per metric family.
  • Documents lane. BM25 + vector search in DuckDB, reciprocal-rank fusion, an optional reranker, and quarantined extraction with exact quotes.
  • Web lane (opt-in). A self-hosted SearXNG, sanitized queries, connect-time SSRF blocking, and external benchmarks that are labelled and can never replace your numbers.
  • Agent. EXPLAIN with fixed tools and causal grades, DECIDE with scenarios, follow-up questions answered with verified claims, and small-model mode.
  • Verifier. V1–V8. Facts are recomputed and headline metrics re-derived, so a mismatch fails the job. Citations are checked, with a judge model. Causal wording is policed. Recommendations must be grounded.
  • Security.
    • Injection screening (EN/RU/UZ, look-alike letters, hidden text) and a quarantine; a 23-fixture injection suite in CI.
    • No free code execution.
    • A DNS-rebinding guard and bearer auth on the API.
  • Doors.
    • CLI: analyze, ask, explain, serve, mcp, jobs, eval, doctor.
    • HTTP API with SSE and OpenAPI.
    • MCP over stdio and HTTP.
    • Telegram bot (long polling).
    • Vite viewer.
  • Report. Self-contained report.html with clickable citations, cited evidence excerpts and a fact panel for every number.
  • Evals.
    • 30 generated businesses with planted events, distractors and quality issues, scored honestly; one-command runs.
    • Smoke results for three free Groq models.
    • The v0.1 bars are not all met yet: see evals/RESULTS.md.
  • Models. Ollama and OpenAI-compatible free endpoints, with rate-limit waits, per-request timeouts and retries of the endpoint’s own JSON validation.