SQLucent

See through any SQL. Paste a gnarly nested-CTE query and get back a data-flow diagram, per-CTE responsibilities, and a plain-language walkthrough — without running the query.

EXPLAIN tells you how the database runs a query. SQLucent tells you what the query means: where data comes from, how it flows through each CTE, and what each step does. It really parses the SQL (via sqlglot), so the structure and diagram are trustworthy — an LLM is only used (later) to polish the prose, never to figure out the SQL.

Status: data-flow graph + walkthrough, interactive HTML, column-level lineage, risk lint, and optional local-LLM narration. Handles SELECT, INSERT, CREATE ... AS, DELETE, UPDATE, and MERGE (multi-statement scripts and Jinja templating included). BigQuery dialect by default. See DESIGN.md for the full design and roadmap.

Demo

Given a nested-CTE query (examples/top_users.sql), sqlucent query.sql --mermaid produces this data-flow graph (rendered live on GitHub):

flowchart TD
  t_events[("events")]
  t_orders[("orders")]
  t_users[("users")]
  n_active["active<br/><i>filter + group-by + aggregate</i>"]
  n_paid["paid<br/><i>filter + group-by + aggregate + having</i>"]
  n_ranked["ranked<br/><i>join + window</i>"]
  n_result(["result<br/><i>join + filter</i>"])
  t_events --> n_active
  t_orders --> n_paid
  t_users --> n_ranked
  n_paid --> n_ranked
  n_ranked --> n_result
  n_active --> n_result

Loading

--html turns the same graph into a self-contained interactive page: click any node to highlight it and inspect that node's SQL, sources, operations, and outputs.

Install (dev)

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Usage

Two commands are installed: sqlucent and the short alias sxr (identical).

sxr examples/top_users.sql                 # short alias for sqlucent
sqlucent examples/top_users.sql            # walkthrough + Mermaid graph
sqlucent query.sql --mermaid               # just the diagram
sqlucent query.sql --walkthrough           # just the steps
sqlucent query.sql --verbose               # full join ON clauses + column lists
sqlucent query.sql --json                  # the IR, for tooling/CI
sqlucent query.sql --html > xray.html      # self-contained interactive page
sqlucent query.sql --lineage               # column-level lineage (all outputs)
sqlucent query.sql --lineage total         # lineage for one output column
sqlucent query.sql --lint                  # risk checks
sqlucent query.sql --lint --fail-on high   # CI gate: exit non-zero on a high finding
cat query.sql | sqlucent -                 # read from stdin
sqlucent query.sql --dialect postgres      # other dialects
sqlucent ./sql/                            # cross-file table lineage for a whole folder

The Mermaid output renders directly on GitHub, in Markdown, and in Notion.

Interactive HTML (--html)

--html emits a single self-contained page (no server) with the data-flow diagram. Click any node to highlight it and inspect that node's SQL, sources, operations, and outputs in a side panel.

By default Mermaid is inlined, so the page is fully offline — open it on a plane, archive it, email it. That makes the file ~3 MB. Use --cdn to load Mermaid from a CDN instead for a tiny (~8 KB) file that needs network to view.

sqlucent query.sql --html > xray.html && open xray.html   # offline, ~3 MB
sqlucent query.sql --html --cdn > xray.html               # tiny, needs network
sqlucent query.sql --html --lang Chinese > xray.html      # localized UI chrome

--lang localizes the page's UI labels (English and Chinese built in; unknown languages fall back to English). SQL identifiers are never translated.

Column-level lineage (--lineage)

Trace each final output column back to the source columns it comes from, followed through every CTE and subquery (powered by sqlglot's lineage engine — no schema required for explicit column references):

sqlucent examples/top_users.sql --lineage
#   user_id     ←  orders.user_id
#   total       ←  orders.amount
#   last_login  ←  events.login_at

Pass a column name to trace just one (--lineage total). SELECT * and bare expressions can't be resolved by name unless you supply a schema:

# --schema takes DDL (CREATE TABLE ...) or a {table: {column: type}} JSON file.
sqlucent "SELECT * FROM users u JOIN orders o ON u.user_id=o.user_id" \
  --lineage --schema examples/schema.sql
#   name    ←  users.name
#   email   ←  users.email
#   amount  ←  orders.amount
#   ...

With a schema, SELECT * is expanded into its real columns and each is traced to its base table (aliases are resolved back to table names).

Project-level table lineage (point at a folder)

Give sqlucent a directory and it scans every .sql file and builds a cross-file, table-level data-flow DAG — which tables feed which, across an entire Airflow/dbt SQL repo, no warehouse connection needed:

sqlucent ./sql/                 # summary + Mermaid table DAG
sqlucent ./sql/ --mermaid       # just the diagram
sqlucent ./sql/ --json          # edges/roots/sinks for tooling

It classifies each table as a source input (read but never written), intermediate (both), or terminal output (written but never read), so you instantly see the roots and leaves of your pipeline. Edges are derived per statement (a statement's source tables → its write target) across INSERT, CREATE, UPDATE, MERGE, and DELETE.

Impact analysis — "if I change this, what breaks?" — walks that DAG:

sqlucent ./sql/ --impact dim_users          # tables downstream of dim_users
sqlucent ./sql/ --impact dim_users.email     # writes that read dim_users.email (best-effort)

Table impact is exact (direct consumers + full transitive downstream, with the first-hop files). Column impact is best-effort without a schema: it flags writes that reference the column explicitly or pull it via SELECT *.

Risk lint (--lint)

Deterministic anti-pattern checks over the AST — no LLM, so findings are stable and CI-friendly:

rule	severity	what it catches
full-table-write	high	UPDATE/DELETE with no WHERE — affects every row
cartesian-join	high	condition-less join between tables (comma/CROSS JOIN), row explosion
select-star	medium	SELECT * — brittle to schema changes, wider scans
having-without-aggregate	medium	HAVING with no aggregate that belongs in WHERE
distinct-with-group-by	low	redundant DISTINCT alongside GROUP BY
partition-filter-missing	high	scan of a declared partitioned table with no WHERE on its partition column — full-table scan, expensive in BigQuery (config-gated)

sqlucent query.sql --lint                  # report findings
sqlucent query.sql --lint --fail-on high   # exit 1 if any finding ≥ high (CI gate)

Config & cost lint (.sqlucent.toml)

Drop a .sqlucent.toml at your project root (auto-discovered from the file's directory upward) to tune rules and declare partitioned tables so the BigQuery cost rule can fire. See examples/.sqlucent.toml:

[rules]
disable = ["distinct-with-group-by"]   # turn rules off

[rules.severity]
select-star = "low"                     # override a rule's severity

[cost]
require_partition_filter = true
[cost.partitions]                       # table -> partition column
events = "event_date"
orders = "order_ts"

Baseline (adopt the linter on a legacy codebase — grandfather existing findings, gate only new ones):

sqlucent query.sql --lint --write-baseline .sqlucent-baseline.json   # snapshot today's findings
sqlucent query.sql --lint --baseline .sqlucent-baseline.json --fail-on high   # only new highs fail
sqlucent query.sql --lint --config path/to/.sqlucent.toml            # explicit config path

Scan-cost estimate (--cost, needs --schema)

BigQuery bills by bytes scanned = referenced columns × rows scanned. Feed a schema (column types) and --cost turns the boolean "missing filter" check into a real number: how much of each row you actually scan (column pruning) and — with row counts from [cost.table_rows] — absolute bytes and dollars.

sqlucent query.sql --cost --schema schema.sql
#   orders: 2/4 cols, 16 B/row (50% of 32 B full row) × 80,000,000,000 rows = 1.2 TB
#   TOTAL ≈ 1.2 TB scanned  →  $8.00 at $6.25/TB

Add a WHERE on the partition column and the same query collapses to a fraction (partition_selectivity), making the savings concrete:

#   orders: 3/4 cols, 24 B/row × 80,000,000,000 rows (partition-pruned) = 35.8 GB
#   TOTAL ≈ 35.8 GB scanned  →  $0.24 at $6.25/TB

Columns referenced only in WHERE/JOIN count (they're read too); COUNT(*) is metadata-only (0 bytes). Variable-width types use the configurable string_bytes assumption, so totals are estimates, not invoices.

Plain-language narration (local LLM, optional)

With Ollama running, --narrate turns the parsed facts into fluent prose. The model only rephrases a deterministic fact sheet built from the parsed IR — it never sees the raw SQL and is told never to invent tables or columns — so the explanation stays grounded. If Ollama isn't running or the model isn't pulled, it falls back to the template walkthrough.

sqlucent query.sql --narrate                      # default model (llama3.2)
sqlucent query.sql --narrate --model qwen2.5:3b   # pick a model
sqlucent query.sql --lang Chinese                 # narrate in any language (implies --narrate)
export SXR_OLLAMA_MODEL=qwen2.5:3b                 # or set a default model
export SXR_LANG=Chinese                            # or set a default language

--lang writes the prose in the given language while keeping SQL identifiers (table/column/step names) unchanged.

How it works

SQL → sqlglot AST → Semantic Model (IR) ─┬─ CTE responsibilities
                                          ├─ data-flow graph (Mermaid)
                                          └─ plain-language walkthrough

The IR is the hub; every feature is a consumer of it. This keeps the deterministic core (parsing, graph, lineage) separate from the probabilistic shell (LLM narration, coming in v0.2).

Releasing

Publishing to PyPI uses Trusted Publishing (OIDC) — no API token or stored secret. One-time setup on PyPI, then every GitHub Release publishes automatically:

1. On pypi.org → your account → Publishing → add a pending trusted publisher:
   • PyPI Project Name: sqlucent
   • Owner: thehwang, Repository: sql-x-ray (the GitHub repo name is unchanged)
   • Workflow name: publish.yml
   • Environment name: pypi
2. In the GitHub repo, create an environment named pypi (Settings → Environments).
3. Cut a release — tag it vX.Y.Z and publish via the GitHub UI or:

gh release create v0.1.0 --title "v0.1.0" --notes "First release."

The publish.yml workflow builds the sdist + wheel and uploads them to PyPI.

License

MIT