decisionbox-platform

DecisionBox is an open-source autonomous AI data discovery platform for data-driven decision-making teams. It connects to your data warehouse and runs autonomous AI agents that decide for themselves what to investigate, write and execute SQL, and ultimately output validated insights and actionable recommendations—all without you having to ask any questions.

Purpose

DecisionBox is for any team that makes decisions based on data. It is not a copilot waiting for questions, but an autonomously operating agent: each discovery run executes 50–100+ SQL queries, entirely unattended.

Key Capabilities

• Autonomous agent loop: AI writes and executes SQL, iterates based on results, and produces insights and ranked recommendations without prompts or questions.
• Insight validation: every conclusion is re-checked in your warehouse via independent validation queries before delivery, avoiding hallucinations and unchecked numbers.
• Full transparency: every SQL query, every reasoning step, and every decision is logged and visible, so you can audit the agent like reviewing an analyst.
• Real-time progress: watch the agent's exploration in real time through phase tracking, step details, and expandable SQL.
• Cost estimation: estimate LLM token and warehouse query costs before running.
• Selective discovery: run specific analysis domains on demand.
• Feedback loop: like/dislike insights and recommendations, and the agent learns from feedback on the next run.
• Semantic search: vector search based on Qdrant + HNSW, covering insights and recommendations, with cross-project scope.
• Ask Insights: RAG-based Q&A providing citations across discovery results, multi-turn sessions, and 90-day search history.
• Spotlight (Cmd+K): quickly navigate and search projects anywhere in the dashboard.
• Knowledge source hooks: connect your own retrieval (documents, runbooks, policies) to enrich exploration, analysis, and /ask prompts.
• Bookmark lists: save insights and recommendations to named lists and add inline notes.
• Read tracking: per-user read status across devices, supporting reduced-opacity rows and a "mark unread" action.
• Technical details toggle: SQL queries and exploration steps are collapsed by default, with one-click expansion of engine internals.
• Related items sidebar: a sticky right-hand table of contents containing related recommendations, related insights, and semantic matches.
• Recommendations and insights cross-linking: every recommendation carries the insight IDs it is based on.

How to Use

1. Connect a data warehouse (BigQuery, Redshift, Snowflake, PostgreSQL, Databricks, Microsoft SQL Server).
2. Configure the project (domain pack, configuration file, LLM provider).
3. Run discovery—the AI agent autonomously explores the data.
4. Review insights—findings sorted by severity, with confidence scores.
5. Act on recommendations—specific, numbered action steps.

Prerequisites: Docker and Docker Compose. The dashboard will guide you through creating your first project, which requires data warehouse connection information and an LLM API key (Anthropic, OpenAI, or configure Vertex AI / Bedrock / Azure AI Foundry). Running locally without Docker is also supported (recommended for development), requiring building the binary.

When to Use It

DecisionBox is applicable to any queryable data. Examples include:

• E-commerce: Cart abandonment spikes 40% when shipping costs exceed 8% of cart value; setting a $75 free shipping threshold is projected to recover 1,200 orders per month.
• SaaS: Teams that do not use dashboard features within 14 days of signup have a 3x higher churn rate; sending an onboarding email on day 3 emphasizing the dashboard can improve activation.
• Fraud detection: Accounts created in the past 48 hours with 5+ high-value transactions account for 82% of chargebacks; flagging this pattern can avoid $34,000 in losses per month.
• Social networks: Posts with images published between 6–8 PM have 3.2x higher share counts, but only 12% of creators post in this window; scheduling reminders can increase platform-wide engagement.
• SQL performance: The 10 slowest queries consume 62% of warehouse compute, 7 of which are full table scans; partition filtering can reduce costs by about $4,800 per month.
• Gaming: Players who fail level 12 more than 3 times have a 68% higher churn rate on day 7; consider adding a hint system or difficulty adjustment at this stage.

Domain Packs

Built-in domain packs: e-commerce, social, gaming. Each domain includes subcategories, so that the agent's analysis areas, prompts, and configuration file schemas match the actual behavior of specific product types. For example, e-commerce covers multiple subcategories because the funnels, retention curves, and profit structures of fashion marketplaces, grocery stores, and digital goods stores are fundamentally different; gaming covers match-3, idle/incremental, and casual/hyper-casual. New domain packs (SaaS, fintech, content/media, education, health, etc.) are released weekly. SupportsDynamic packs: create, edit, import, export from the dashboard (stored in MongoDB, no code changes needed), portable JSON format shareable across projects and environments.

Supported warehouses and LLMs

Warehouses: BigQuery (ADC or service account key, supports dry-run cost estimation), Redshift (Serverless + provisioned; IAM role, access keys, or Assume Role cross-account with external ID), Snowflake (username/password or key pair JWT), PostgreSQL (username/password or connection string, SSL configurable), Databricks (Unity Catalog via personal access token or OAuth M2M service principal), Microsoft SQL Server (SQL Server 2016+ and Azure SQL Database, SQL login or full connection string). Each provider ships with SQL repair prompts tuned to its dialect and enforces read-only and schema-aware SQL self-healing. More warehouses ship weekly. LLMs: Claude (direct API), OpenAI, Ollama (local), Vertex AI (Claude + Gemini on GCP), AWS Bedrock, Azure AI Foundry, LiteLLM (self-hosted OpenAI-compatible proxy). Supports per-project configuration, editable prompts, per-model max output token limits. Custom TLS (private CA upload or skip verification) for local HTTPS endpoints. Embedding providers: OpenAI, Vertex AI, Bedrock, Azure OpenAI, Voyage AI, Ollama, for insight/suggestion embeddings, semantic search, and /ask.

Deployment and scaling

• Docker Compose for local development single-server installation.
• Helm charts (API + Dashboard + optional MongoDB + Qdrant sub-charts), public chart repository see https://decisionbox-io.github.io/decisionbox-platform。
• Terraform modules support GCP (GKE), AWS (EKS), Azure (AKS), with Workload Identity / IRSA wiring.
• Interactive setup wizard (terraform/setup.sh) supports GCP, AWS, or Azure, with authentication, restore, and destroy.
• Optional IP allow list for GKE / EKS / AKS control planes and dashboard ingress.
• Multi-architecture Docker images (linux/amd64 + linux/arm64).
• Plugin architecture: providers register via init() + RegisterWithMeta(); warehouse middleware (warehouse.RegisterMiddleware()) can wrap provider implementations for logging, metrics, or access control; HTTP middleware (apiserver.RegisterGlobalMiddleware()) can wrap all API requests (audit logging, custom authentication); custom builds via apiserver.Run() / agentserver.Run() and blank imports.

Credentials and notifications

Per-project credentials stored encrypted, supporting MongoDB (default AES-256-GCM), GCP Secret Manager, AWS Secrets Manager, Azure Key Vault. Webhook notifications on discovery completion: Slack, generic HTTP, or email, with per-project configurable templated payloads.

Limitations and notes

• Requires bringing your own data warehouse connection and LLM API keys.
• Licensed under AGPL v3, self-hostable.
• Collects anonymous usage telemetry (no PII, query content, or credentials), can be disabled via TELEMETRY_ENABLED=false or DO_NOT_TRACK=1.
• For specific installation commands, the full list of environment variables, and configuration details, refer to the official documentation at decisionbox.io/docs.