Semantic Explorer -- DuckDB WASM Explorer

Drop data files here or click to browse CSV, TSV, Parquet (recommended in browser), JSON, JSON Lines, Arrow IPC, DuckDB (.duckdb).

Privacy Local execution

Privacy & Local Execution

Semantic Explorer runs all analysis locally in your browser using DuckDB WASM. Your data is loaded into browser memory and is not uploaded to any server.

You can verify this by disconnecting from the internet after the app loads and continuing to query the data.

Engine: DuckDB WASM
Execution: Browser (local)
Data upload: None (client-side processing)

Joins Analysis

Composite Key Analysis

Column anomaly detection

Convert CSV to Parquet

Scan mode

When a table has more than 20,000 rows (from profiling), hybrid keeps distributions on a 10k sample but runs full-table COUNT and COUNT DISTINCT per column for keys and joins. Sample uses the same 10k reservoir for all column metrics and for join validation.

Pre-load WebLLM for SQL assistance (requires WebGPU; about 2-2.5 GB browser cache for large models)

Model

Summary

Analysis details (full tables)

Tables

Data Modeling & Structure

Data Cleaning & Normalization

Join Safety & Corrections

Join Gap Diagnostics

Column Profiles

Detected Keys

Join Candidates

Recommended Joins

▶ Join Validation (all tested pairs)

1Base table and joins

▶ Joins

Base table (suggested from FK structure)

Available joins

Custom join

Add a join by choosing tables and columns (included in generated SQL when checked).

Left

Right

▶ Join safety alerts

▶ Entity Relationship Diagram

Max rows

Saved Explorations

Each saved exploration stores your comment, the SQL, tables and join references inferred from that query (not the whole sidebar), and column statistics from the query result only (not source-table profiling). Download YAML to use as AI context (same schema as explorations.yaml in Context Download).

Context Download

Download markdown and YAML artifacts from the current analysis session. Use them as context in Cursor, Claude Code, or any other AI assistant.

About these downloads

The bundle is context.md plus YAML exports. Attach context.md by default; add YAML only when you need exact metrics or full join rows. Do not treat profiling-derived output as production-ready without business review.

Files

context.md — Short guidance for AI readers: inferred grain, table type, composite join candidates, and similar labels are heuristic, not business-approved. The body after that guidance is intentionally empty until you run Generate AI context summary, which appends a deterministic markdown digest (profiling highlights, promoted joins, cleaning, join safety, anomalies, saved exploration summaries including detached-pivot Perspective layout hints) and points readers to the YAML files for full detail.
analysis_data.yaml — Full metrics: profiling, cleaning, all join candidates and validations, promoted_safe_recommended vs rejected_or_unsafe_recommended (plus session-all list), safe_modeling summary (dbt test hints, exploration join warnings), composite_keys, join safety, anomaly_detection.by_column. Attach only when exact numbers matter.
explorations.yaml — Saved notes from the Query Explorer or detached pivot: comment, SQL, tables and joins, optional result_column_profiles, perspective_viewer_save for pivot saves, and join_warning when saved join edges are not in the promoted safe set.

Using them in an AI assistant

Cursor / Claude Code. Attach context.md by default. Attach analysis_data.yaml only when you need exact metrics or full join detail. Attach explorations.yaml only when analyst comments or saved SQL materially matter. Do not blindly implement every profiling-derived suggestion.

Claude Code and other tools. Same order: narrative guardrails, then structured YAML. Regenerate the optional deterministic summary in the app if the session changed, then download again.

Run an analysis first to generate downloadable files.