Natural Language BI Platform

Status: Candidate Project
Market Size: $34.8B (2025) → $72.2B (2034) at 8.4% CAGR (BI market); Augmented Analytics growing at 17.4% CAGR
Last Updated: 2026-05-02

Overview

An open-source, self-hostable natural language business intelligence platform that treats semantic accuracy as a first-class citizen. Today's NL BI tools (ThoughtSpot, Power BI Copilot, Tableau Einstein) struggle with complex schemas and ambiguous metrics. This project solves accuracy through:

Mandatory semantic layer grounding: Accuracy improves from 70–85% baseline to 90%+ through Cube.js or dbt Semantic Layer integration
Conversational drill-down with context retention: "Now break that down by region"—without re-specifying the full query
Automated insight narration: LLM-generated written narratives alongside charts explaining what changed and why it's notable
Self-healing dashboards: Auto-detect schema changes and propose semantic-similarity-based column mappings
Open-source + permissive license: Escape vendor lock-in; modify and white-label freely

The Market Gap

The BI market is bifurcated:

Enterprise vendors ($25–$200/user/month): ThoughtSpot, Power BI Copilot, Tableau Einstein—mature but expensive and vendor-locked
Open-source BI (free): Metabase, Superset—great for dashboards but zero NL querying out of the box
Emerging AI-native SaaS ($20–$5,000/mo): Julius AI, camelAI, Supaboard—no semantic layer, high hallucination rates, expensive at scale

The gap: No open-source tool combines production-grade NL querying with semantic grounding. The Cube.js + Metabase combination is technically possible but requires substantial engineering. Teams want:

Free or low-cost entry (eliminates $100K+/yr contracts)
NL query accuracy on real enterprise schemas (hallucination is a blocker)
Self-hosting for data sovereignty
Permissive licensing for white-labeling

Core Features

MVP (Must-Have)

Natural language question interface producing a chart or table result, grounded in a semantic layer (Cube.js or dbt Semantic Layer) to anchor accuracy and prevent hallucination
Connectors for at least PostgreSQL, MySQL, BigQuery, and Snowflake as initial data sources
Multi-turn conversational context: follow-up questions that reference and refine the previous result without re-specifying the full query
Role-based access control ensuring users can only query data within their permission scope
Dashboard creation and shareable link generation from NL-derived visualisations

Should-Have (v1.1)

Automated insight narration: LLM-generated written summary of what the data shows, what changed, and what is notable—alongside each visualisation
Schema-aware query accuracy improvements: RAG over schema documentation and metric definitions; query decomposition for multi-join questions
Anomaly and trend alerts: scheduled monitoring of key metrics with NL narrative alerts delivered via Slack or email
Embeddable NL query widget with signed-URL embedding API for SaaS product integration

Nice-to-Have (Backlog)

Self-healing dashboards: Detect broken column/table references after schema migrations and propose semantic-similarity-based fixes
Proactive insight delivery: push weekly metric summaries to users without requiring them to log in
NL-driven semantic model authoring: allow data engineers to describe a metric in plain language and have the system generate Cube.js or dbt YAML definition
Export to presentation formats (PDF, slides) with LLM-generated narrative captions per slide

AI-Native Opportunities

Accuracy through semantic grounding
- Current NL-to-SQL tools achieve 70–85% accuracy on clean-schema benchmarks; real enterprise schemas are messier
- A platform with mandatory semantic layer integration (dbt Semantic Layer / Cube) can push accuracy to 90%+ by constraining LLM output to governed metrics only
- This closes the gap that makes current tools untrustworthy for business decisions
Conversational drill-down and context retention
- Existing tools treat each query as stateless
- An AI-native platform maintaining session context enables: "now break that down by region," "exclude the Q2 outlier," "compare to same period last year"—reducing time-to-insight from hours to seconds
Automated insight narration
- Most BI platforms produce charts that users must interpret
- An LLM layer generates written narratives alongside visualizations—explaining what changed, why it's notable (anomaly/trend), and what action it implies—making dashboards accessible to non-analysts
Self-healing and schema-aware query repair
- When schema changes break dashboards (renamed columns, deprecated tables), current systems fail silently or require manual fixes
- An AI agent could detect failures, map old references to new ones using semantic similarity, and propose or auto-apply fixes
Open-source differentiation
- Power BI Copilot, Tableau Einstein, ThoughtSpot are tightly coupled to vendor ecosystems and pricing
- An open-source platform with pluggable connectors (any SQL source), permissive license, and community-governed semantic layer would be the first true open-source NL BI alternative

Competitive Landscape

Tool	Type	Semantic Layer	Accuracy	Cost	Self-Hosted
ThoughtSpot (Spotter)	Commercial SaaS	✓ (Spotter Semantics)	85%+	$25–$300K/yr	Limited
Power BI Copilot	Commercial SaaS	Limited	75%	$14/user/mo + $262/mo Fabric	❌
Tableau Einstein	Commercial SaaS	Partial	80%	$75+/user/mo + premium	❌
Looker (Gemini)	Commercial SaaS	✓ (LookML)	85%+	$30K–$150K/yr	Limited
Metabase	OSS + SaaS	❌	40% (without Semantic Layer)	Free self-hosted	✓
Cube.js	OSS semantic layer	✓	N/A (engine-agnostic)	Free self-hosted	✓
This Project	OSS + SaaS	✓ (required)	90%+	Free self-hosted	✓

Technical Design Considerations

Architecture: Semantic layer integration required; support Cube.js (REST/GraphQL/SQL APIs) and dbt Semantic Layer (MetricFlow)
NL to SQL: LLM (Claude, OpenAI) with RAG over semantic model definitions; query decomposition for multi-join questions (DIN-SQL approach)
Session state: In-memory context management per user; optional Redis for distributed deployments
Visualization: Embed or build on Apache Superset's 40+ chart types
Accuracy grounding: Semantic layer constraints prevent hallucination; LLM output must conform to defined metrics
Deployment: Docker Compose or Kubernetes; managed SaaS option with team collaboration

Market Validation

Market drivers:
- Augmented analytics (AI-powered BI) growing at 17.4% CAGR—fastest-growing sub-segment
- Citizen analyst demand: business users need ad-hoc querying without data team dependency
- Vendor lock-in fatigue: enterprises evaluating open-source alternatives to Power BI/Tableau/Looker
Customer personas:
- Business analysts at mid-to-large enterprises without dedicated SQL skills
- Data-forward SMBs (50–500 employees) that cannot afford $100K/yr contracts
- Product managers and C-suite executives wanting ad-hoc query access
- Embedded analytics buyers (SaaS companies white-labeling NL BI)

Why Build This

Market timing: Augmented analytics growing at 17.4% CAGR; open-source gap is wide
Technology maturity: Text-to-SQL via LLMs is well-researched; semantic layer standards (Cube, dbt) are established
Commercial opportunity: Build on free tier; offer managed SaaS for enterprises (similar to Metabase model)
Platform leverage: Cube.js provides semantic layer; Superset provides UI; Claude API adds conversational layer

Success Metrics

Adoption: 1K+ GitHub stars within 12 months; featured as Metabase alternative for NL querying
Accuracy: Achieve 90%+ query accuracy on real enterprise schemas (vs. 70–85% baseline)
Commercial: Win 10+ SMB/mid-market customers with managed SaaS tier
Enterprise: Demonstrate ROI (reduce time-to-insight from hours to minutes) with 5+ case studies

Resources:

claude