What's new in DecisionBox since April: v0.5 to v0.37
Since v0.4.0 in April we have shipped 39 open-source releases, and this post brings them together in one place. The biggest changes: Ask writes and runs its own SQL, two AI agents check findings against your data, and the Discovery Ledger carries what a run learned into the next one.
Ask runs its own queries
Our April post about Ask described answers drawn from findings DecisionBox had already made. Ask now answers new questions too. It writes read-only SQL, which reads your data without changing it, runs several queries against your warehouse, and answers from the results. When the data can't support an answer, it tells you. Charts are drawn from the same query results.
You can also open Ask from any finding and keep digging from there.
Edition: answers that run new SQL, and charts, are in Enterprise. Asking about past findings is in the Community edition (open source). More on the Ask page.
How does DecisionBox check a finding?
Validation uses two AI agents. One tries to confirm the finding against your data. The other tries to prove it wrong. A finding that fails is labeled, so you can see what was tested and what didn't hold up. You can also run validation on demand for a single finding, or switch it off for a project.
Claim rules also tell the agents to stay inside the data that was actually queried, so a run that looked at 90 days shouldn't call something an all-time record.
Edition: Community edition. The validation docs explain the checks and their limits, and episode 3, Every claim cross-validated, shows them on a real run.
Discovery that builds on the last run
The Discovery Ledger carries findings, open questions and lessons into the next run, so each run continues the investigation instead of repeating it. When DecisionBox isn't sure what something in your data means, it asks you, and your answer becomes part of what it knows. Every finding traces back to the SQL and the rows behind it.
Edition: Enterprise. More on the Discovery Ledger page and in episode 11, It remembers.
From a conversation to a file you can send
After every run, DecisionBox writes an executive summary with a citation for every figure. An Ask conversation can also become a PDF, Excel, Word or PowerPoint file. You change the file by asking, and its numbers come from query results. The model never types them in.
Edition: Enterprise. Episode 6, Cited answers, instant reports, shows both.
More places your data can live
SQL Server joined the Community edition, which now connects to six SQL warehouses: BigQuery, Redshift, Snowflake, PostgreSQL, Databricks and SQL Server. Enterprise adds Oracle and SAP HANA, several warehouses in one project with Ask picking the right one for each question, and live queries against Google Analytics 4.
The provider docs list the Community connections, and episode 1, From warehouse to first findings, shows a warehouse being connected.
A model for each step
DecisionBox works with seven language-model providers: Claude, OpenAI, Ollama, Vertex AI, AWS Bedrock, Azure AI Foundry and LiteLLM. That includes model endpoints inside your own network that use private certificates.
Since v0.37 you can choose a provider and model for each place DecisionBox calls a model, including exploration, analysis, the two validation agents and Ask. A cheaper model can do the exploring while your strongest model does the verifying.
Edition: Community edition. Setup is in the v0.37.1 configuration guide. Episode 13: Your model, your rules.
A Claude run cost 76% less in our test
v0.36 added prompt caching, which lets the model reuse the parts of a request it has already seen, so those parts cost much less the second time. We compared runs with the same model and settings on the same day. Exploration cost fell by 95% and the whole run by 76%, and the release notes report the same quality profile.
Edition: Community edition. The measurements are in the v0.36.0 release notes.
Access controls and release verification
For teams that self-host, Enterprise adds custom roles with per-project access, OpenShift deployment support, offline installs, scheduled runs and results by email.
Signed images and signed release tags are in the Community edition too, and the verification guide shows how to check them. Episode 14 covers governance and audit.
Smaller changes
- Findings, recommendations and Ask answers can be written in 20 languages.
- DecisionBox recognizes 18 more open model families when you run models locally with Ollama.
- Every language-model call records its token use, including cached tokens, and the phase of the run that made it.
Upgrade notes
After upgrading to v0.37, open each existing project's settings once and save its model configuration. The values are already filled in from what the project uses today, so it's a quick review. Until then, discovery won't start on that project and its scheduled runs stop, with the reason recorded in the run history. The configuration guide has the details.
Prompt caching is on by default, and you can switch it off for each model setting in Settings → Models. The release notes have the details for each published release.
Try it
The Community edition is on GitHub, and the self-host guide takes you to a first run.