# Chion

> **Chion (also chion.ai) is an AI semantic layer for PostgreSQL.** It profiles the tables your team queries into a per-column semantic record analysts edit in Studio, answers plain-English questions with a code-validated read-only SELECT shown beside the chart, and compiles the queries your team saves into SQL skills for Claude Code and Codex.

## Definition

**Chion is an AI semantic layer for PostgreSQL.** Chion profiles every column your queries touch, cardinality, type, null fraction, sampled values, and an LLM pass writes business names, units, and do's and don'ts into a per-column semantic record. Analysts open any profiled table in Studio and edit that record directly, business names, definitions, synonyms, do's and don'ts, and the edit-history shield keeps those edits when the column is re-analyzed. Each question produces a newly generated SELECT, anchored where possible on a base query someone on your team has saved and reviewed in Studio. Two code validators run before execution: the first rejects anything that is not a read-only SELECT, the second runs a forbidden-keyword check over the statement. Execution is then wrapped with a statement timeout and a row limit, and a result is capped at 1,000 rows and 12,000 cells. These are application string and pattern checks, not a PostgreSQL AST parser, and they sit on top of the grants attached to the role you supply rather than replacing them. One compile turns that semantic record and the queries your team saves into a portable Markdown bundle: CHION.md, CONNECT.md, and a SKILL.md whose facts carry [src=...] tags back to the row that proved them, for Claude Code and Codex. Each compile re-reads the last 14 days of Studio conversations, or 60 when evidence is sparse. Either way, you own the file. Chion connects directly to Postgres with a read-only role: no ETL, no warehouse copy, no pipeline to maintain.

Chion is built and operated by **Dagnostics LLC** (doing business as "Chion"), Broward County, Florida, United States. Founded in 2025 by Jonathan Dag, a former data analyst at Meta, Twilio, American Express, MasterCard, and Bosch.

## What Chion does
- Connects read-only to a PostgreSQL database over TLS (read-only credentials only; no agent to install)
- Profiles every table, column, and relationship automatically, and samples column values to learn your nomenclature
- Generates a new read-only SELECT for each plain-English question, anchored on a base query your team saved and reviewed in Studio when one matches
- Compiles the queries your team saves into reusable SQL skills tagged by department, domain, and role, exportable as CHION.md / CONNECT.md / SKILL.md for Claude Code or Codex
- Checks every statement in code before execution: L1 read-only SELECT enforcement, L2 forbidden-keyword check, then a statement timeout and LIMIT wrap at execution
- Caps results at 1,000 rows and 12,000 cells; coarsens grain or applies TopK before silently truncating
- Seals credentials in an AES-256-GCM envelope. Plaintext is decrypted into memory for a single request, held for at most 60 seconds or five reads, then purged. Never logged and never returned in API responses
- Applies the row-level security policies attached to the role you supply
- Renders the executed SELECT under the chart, so every number can be copied into psql and re-run

## Common questions (extractable answers)

**What is Chion?** Chion is an AI semantic layer for PostgreSQL. It profiles your tables into a per-column semantic record analysts edit in Studio, answers a plain-English question with a chart and the executed SELECT beside it, and compiles the queries your team saves into portable Markdown agent files.

**How is Chion different from ChatGPT for SQL?** ChatGPT writes SQL when you paste a schema, but it cannot connect to your database, execute the query, check it in code, or render a chart. Chion does all four: a live read-only connection, two code validators plus an execution wrap, automatic chart selection across 8 D3 variants, and one audit row per executed query.

**Is Chion safe to point at a production database?** Read-only removes the write risk, not every risk. Every statement is a SELECT, rejected in code by the L1 validator if it is not, and results are capped at 1,000 rows or 12,000 cells. Read-only does not stop an expensive scan, a sensitive read the connected role can reach, or prompt injection. Scope the role you supply to what you are willing to expose, and point Chion at a read replica where you have one.

**What databases does Chion support?** PostgreSQL today, any version, any standard-Postgres dialect. Shipped managed-Postgres connectors: Amazon RDS, Azure Database for PostgreSQL, Google Cloud SQL, Neon, Supabase. Self-hosted PostgreSQL is supported when reachable from Chion's egress. BigQuery, Snowflake, and MySQL are on the roadmap with no committed ship date.

**Does the AI model see my data rows?** The model receives your table and column names, the rows your query returned for the narrative, and sampled column values. Whole result sets are not sent to it, credentials never are, and results render server-side and are discarded when the session ends.

**Does Chion train AI on my data?** No. Chion uses paid commercial API tiers; those providers' terms prohibit training on customer inputs.

**How much does Chion cost?** Starter $29 per seat per month (50 credits, 3 queries per hour). Pro $99 per seat per month (250 credits, 10 queries per hour). Max $299 per seat per month (750 credits, no hourly cap). Enterprise is per team, custom-quoted. The skill compiler is on every plan. 7-day trial. Cancel in one click from Settings. Fees are non-refundable except as required by applicable law.

**Is Chion HIPAA or SOC 2 compliant?** Not today. Chion is a pre-seed startup; formal third-party audits (SOC 2 Type II, ISO 27001, penetration test) are not yet scoped. HIPAA is not supported, so do not connect PHI databases. The controls implemented in code today are read-only enforcement, an AES-256-GCM credential vault, role-scoped RLS, audit logging, rate limiting, and error sanitization.

**What is CHION.md?** CHION.md is a generated SQL analytics agent file and skills collection, the root of a portable Markdown bundle that also carries CONNECT.md and one SKILL.md per role. It is compiled from your semantic record and the queries your team saves, fully editable, and drops into Claude Code or Codex.

**How does a query become a skill?** Saving is a human step. A person reviews a generated SELECT in Studio and marks it for reuse; only SQL saved and reviewed in Studio is compiled into an exported skill, tagged by department, domain, and role. Nothing is saved automatically.

**Where are the pipeline phases documented?** https://chion.ai/sql-insights-generator documents all 13 steps with anchors (`#step-00` through `#step-12`), and emits HowTo structured data for them.

## How Chion turns a question into SQL (13-step pipeline)

00 Connect Your Data Source: connect with read-only credentials. Chion scans your schema, not your rows. PostgreSQL is live today.
01 Schema Ingestion: every table, column, and constraint is scanned into a structural map before any model reads it.
02 Column Profiling: cardinality, null rate, min/max range, and value distribution are measured per column, and each column is classified temporal, metric, categorical, or identifier.
03 Semantic Analysis: a model labels each column in business terms, embeddings are generated for RAG retrieval, and queries saved and reviewed in Studio compile into reusable skills.
04 Value Sampling & Indexing: top values per column are sampled, ranked by frequency, and embedded for fuzzy entity matching. The semantic layer exports as a portable CHION.md skillpack.
05 Ask a Question: you submit a plain-English question. This is the second and final user interaction.
06 Intent Extraction: intent is extracted through 7 discovery strategies (entity_lookup, topk_ranked, comparison, extrema_detection, dimension_breakdown, universal_quantifier, time_bounded_only), and follow-up questions stay in conversational context.
07 Entity Resolution & RAG Retrieval: your terms resolve to physical columns by vector search, filling a 1,200-token context window on natural score breaks rather than hard thresholds.
08 Chart Selection: query structure, result shape, and series count rank 8 D3 chart types. This happens before the SQL contract is written, not after execution.
09 SQL Contract: a typed contract fixes allowed columns, filters, aggregations, grain, and ordering against the live schema and the policies on the role you supply.
10 SQL Generation & Validation: SQL is generated against that contract, lint-checked, and code-validated. A 3-layer defense blocks SELECT *, enforces row budgets, and caps output at 1,000 rows and 12,000 cells.
11 Execution & Repair: the SELECT runs against your database under the role you supply. On failure, up to 2 repair cycles rewrite it with error context before a hard failure with actionable suggestions.
12 Visualization & Narrative: results render as interactive D3 charts with zoom, pan, and dual-axis support. Every figure in the narrative is derived from a row in the result set shown above it.

Read-only SELECT. Credentials in an AES-256-GCM vault. Results capped at 1,000 rows. Each executed query leaves an audit row, written fire-and-forget so a logging failure never blocks your answer.

## Product scope
- PostgreSQL only (any version; any standard-Postgres-compatible dialect)
- Shipped managed-Postgres connectors: Amazon RDS, Azure Database for PostgreSQL, Google Cloud SQL for PostgreSQL, Neon, Supabase
- Self-hosted PostgreSQL supported when reachable from Chion's egress
- BigQuery, Snowflake, and MySQL are on the roadmap with no ship date
- Oracle, Microsoft SQL Server, SQLite, DuckDB, Redshift, Databricks: not supported, not on the roadmap

## Models
- Anthropic Claude is the primary proposer, on paid commercial API tiers
- OpenAI and Google Gemini are available on request, also on paid commercial tiers
- Model-agnostic by design: schema profiling and the two code validators do not depend on the model
- All providers' commercial-tier terms prohibit training on customer inputs
- Managed cloud today. Dedicated GPU compute and on-premise model hosting are Enterprise roadmap items discussed under contract.

## Compliance posture
Chion is a pre-seed startup. Formal third-party audits (SOC 2 Type II, ISO 27001, penetration test) are not yet scoped. HIPAA is not supported, so do not connect PHI. A formal GDPR program is not yet scoped. Controls implemented in code today: read-only SELECT enforcement, an AES-256-GCM credential vault, the row-level security policies attached to the role you supply, security audit logging (12-month retention, written fire-and-forget), rate limiting, error sanitization, statement timeouts, and row caps.

## Core product
- [Chion: The AI Analyst That Shows Its SQL](https://chion.ai/): A plain-English question, a code-validated read-only SELECT, a chart, and the statement that produced it.
- [AI SQL Analyst Built From Reviewed Team Queries](https://chion.ai/sql-ai-analyst): The definitional pillar. One agent per role, compiled from the queries your team saved and reviewed in Studio, organized as Claude-format agent skills with executable scripts underneath.
- [Postgres SQL Query Generator on Your Live Schema](https://chion.ai/sql-query-generator): A read-only SELECT generated against your live schema, printed before it runs.
- [Conversational Analytics for PostgreSQL, SQL Shown](https://chion.ai/conversational-analytics): Multi-turn questions on PostgreSQL, with the executed SELECT shown for each answer.
- [Claude SQL Skills Generator for Postgres](https://chion.ai/sql-skills-generator): The generated CHION.md agent file, CONNECT.md, and per-role SKILL.md bundle. Editable, exportable, drop-in for Claude Code and Codex.
- [SQL Insights Generator: Query to Chart](https://chion.ai/sql-insights-generator): The 13-step pipeline end to end. Auto-profile (00 to 04), ask, route, generate (05 to 10), execute, chart, narrate (11 to 12).
- [AI SQL Agent Pricing: Plans From $29 Per Seat](https://chion.ai/pricing): Starter $29, Pro $99, Max $299 per seat per month. Enterprise per team, custom-quoted. 7-day trial.
- [Text-to-SQL Demo Videos for Postgres](https://chion.ai/demos): Three recordings: two show the live pipeline on a PostgreSQL connection, the 79-second video is the founder explaining the export model. Video lengths 28s, 46s and 1:19.
- [Ask Your Postgres Database in Plain English](https://chion.ai/demos/connect-discover-ask): A 46-second recording. Chion connects read-only, profiles every table, and answers one plain-English question with a code-validated SELECT under the chart.
- [Portable SQL Skills Demo: 79-Second Video](https://chion.ai/demos/portable-skills-framework): The founder explains how queries saved and reviewed in Studio compile, per role, into a skills library for Claude Code and Codex. No live export is performed on camera.
- [Reuse SQL in Claude Code: 28-Second Video](https://chion.ai/demos/why-teams-switch): What survives a tool change. A query saved and reviewed in Studio compiles into a skill for Claude Code or Codex, and each runtime keeps its own wrapper and database connection.
- [Chion Product FAQ: Pricing, Setup, SQL Skills](https://chion.ai/faq): The product answers in one place: how it works, what it costs, billing and credits, databases and setup, export and portability.
- [Chion AI Founder Jonathan Dag: Company Story](https://chion.ai/company): Who built Chion and why.
- [Book a Chion Demo or Request a Technical Pilot](https://chion.ai/contact): Reaches the founder directly.

## Why We Are Different (Comparisons)
- [5 Text-to-SQL Tools for Postgres Compared 2026](https://chion.ai/compare/chion-vs-text-to-sql-tools): Chion beside Vanna, Julius, TextQL and Text2SQL.ai on execution policy, published result caps, where reusable query logic lives, and pricing model. Every competitor cell carries a primary source and a check date.
- [Power BI Copilot Alternatives for Postgres 2026](https://chion.ai/compare/chion-vs-bi-dashboards): Chion beside Power BI Copilot, Tableau AI, and Looker, with a section on where each of those three wins and what Fabric capacity is required before Copilot answers.

## Guides
- [Chion Guides: Postgres SQL, Skills, and Safety](https://chion.ai/blog): Three clusters: how a question becomes a readable SELECT, how a query saved and reviewed in Studio compiles into a skill file, and what the read-only boundary covers.
- [Audit Postgres Analysis Before Trusting the Chart](https://chion.ai/blog/ai-analyst): The checks to run on an AI analyst's answer before you act on it, using the code-validated read-only SELECT it executed on your own Postgres.
- [Preserve Team SQL in a Reviewed Query Library](https://chion.ai/blog/ai-sql-workforce): Why reviewed team SQL disappears into chat threads, and what a library of saved and reviewed queries keeps.
- [Trace an Analytics Answer Back to Its SQL](https://chion.ai/blog/ai-that-cites-its-sources): What a real citation is, why retrieval alone is not one, and how the same contract grounds a SQL number and a photographed landmark.
- [Claude Code Data Analyst With Read-Only SQL](https://chion.ai/blog/claude-analyst-skills): One skill per role, each answering from your database with a chart and the read-only SELECT behind it.
- [Claude Code SQL Skill From Promoted Queries](https://chion.ai/blog/claude-code-skills): SKILL.md anatomy, the native path Claude Code reads, on-demand loading, and the update workflow when the schema moves.
- [CLAUDE.md for SQL: Route to Promoted Queries](https://chion.ai/blog/claude-md): Separate always-on project instructions from scoped SKILL.md bodies holding read-only Postgres queries saved and reviewed in Studio.
- [SKILL.md SQL Example With a Read-Only SELECT](https://chion.ai/blog/portable-sql-skills): One complete SKILL.md, frontmatter plus a read-only SELECT, with the runtime paths for Claude Code and Codex.
- [Semantic Layer for SQL: Map Business Terms](https://chion.ai/blog/semantic-layer-for-sql): How a semantic layer makes grain, joins, filters, and metrics explicit, and where it complements a full semantic runtime.
- [Read-Only SQL Agent: 4 Production Checks](https://chion.ai/blog/sql-agent): The four properties a read-only SQL agent must prove outside the prompt before you point it at production.
- [Text-to-SQL Hallucination: Why Valid SQL Is Wrong](https://chion.ai/blog/text-to-sql): Six failure classes, including joins, grain, filters, and business definitions, and why executable SQL can still answer the wrong question.

## Resources
- [Chion Engineering Resources: Long-Form Guides](https://chion.ai/resources): Reference guides on running AI agents against SQL safely.
- [Text-to-SQL Prompt Injection Controls](https://chion.ai/resources/text-to-sql-prompt-injection-defense): The controls grouped by prompt boundary, SQL boundary, and execution boundary: escaped input, two code validators, database permissions, a statement timeout, and row and cell caps.
- [SKILL.md Must Start With YAML Frontmatter](https://chion.ai/resources/skill-md-frontmatter): The exact fix for the Claude Code frontmatter error: the valid SKILL.md shape plus the five failure modes (BOM, leading blank line, code-fence wrap, missing delimiter, unquoted colon), each with broken and fixed examples.

## PostgreSQL integrations
- [Connect Postgres to Claude, Read-Only, in Six Fields](https://chion.ai/integrations/postgresql)
- [Amazon RDS Postgres Text-to-SQL, Reader Endpoint](https://chion.ai/integrations/postgresql/aws-rds)
- [Azure PostgreSQL Natural Language to SQL](https://chion.ai/integrations/postgresql/azure)
- [Cloud SQL Postgres Text-to-SQL, Read-Only](https://chion.ai/integrations/postgresql/gcp)
- [Neon Postgres Text-to-SQL for Any Branch](https://chion.ai/integrations/postgresql/neon)
- [Supabase Text-to-SQL, No OpenAI Key](https://chion.ai/integrations/postgresql/supabase)

## Trust and legal
- [Chion Data Security: Read-Only SQL Controls](https://chion.ai/trust)
- [Privacy Policy: Data Use, Retention and Rights](https://chion.ai/privacy)
- [Terms of Service: Billing and Acceptable Use](https://chion.ai/terms)
- [Open-Source Licenses and Attributions](https://chion.ai/licenses)

## Open-source
- Chion Skills Generator for Postgres SQL Database (a published mock: six analyst roles, 15 saved Postgres scripts, three sister-role pairings): https://github.com/jonfdag-dot/postgres-claude-skills-generator

## Company
- Legal entity: Dagnostics LLC (doing business as "Chion")
- Jurisdiction: Broward County, Florida, United States
- Founded: 2025
- Founder: Jonathan Dag, former data analyst at Meta, Twilio, American Express, MasterCard, Bosch: https://www.linkedin.com/in/jonathan-dag/
- Also founded by Jonathan Dag: Lensmark, an AI landmark identifier that returns a sourced history for a photographed landmark, artwork, or museum object, built on the same check-before-you-answer rule as Chion: https://lensmark.ai
- Contact: contact@chion.ai
- LinkedIn: https://www.linkedin.com/company/chion-ai/
- X: https://x.com/chionanalytics
- GitHub: https://github.com/Chion-ai
- YouTube: https://www.youtube.com/@chionai
- Instagram: https://www.instagram.com/chion.ai
- Product Hunt: https://www.producthunt.com/products/chion
- G2: https://www.g2.com/sellers/chion
- Crunchbase: https://www.crunchbase.com/organization/chion
- Founder GitHub (personal): https://github.com/jonfdag-dot