Forge Discovery Guide
Removed in 0.31.0
(since 0.31.0, forge-cli #881) Removed: local file discovery (--discover, --no-discover, --discovery-path) on fluid forge. The coding agent reads the workspace itself. The text below describes forge-cli 0.29.0.
This guide explains how the current fluid forge flow discovers local context before it scaffolds a project.
Current public entry point
Use fluid forge, not the older fluid forge --mode copilot examples that may still appear in archived material.
Examples:
fluid forge
fluid forge --discovery-path ./data
fluid forge --discovery-path ../shared-schemas
fluid forge --no-discover
What discovery is for
Discovery gives Forge grounded local context so it can:
- ask fewer follow-up questions
- infer useful provider or domain hints
- summarize local schemas without sending raw rows
- generate a stronger first contract draft
What Forge scans
Forge may inspect local:
- SQL files
- dbt projects
- Terraform files
- existing FLUID contracts
- README headings
- sample CSV, JSON, JSONL, Parquet, and Avro files
Optional helpers
These helpers are not the AI backend. Generating a contract with an API-key provider needs the litellm extra (pip install "data-product-forge[litellm]"). (since 0.28.0, forge-cli #837)
If you want richer Parquet and Avro schema inspection, install the optional discovery helpers:
pip install "data-product-forge[copilot]"
That adds local readers such as pyarrow and fastavro. Without them, Forge can still discover files, but schema extraction is more limited.
Privacy boundary
Discovery is metadata-first.
Forge sends distilled summaries such as:
- column names
- inferred scalar types
- referenced tables
- provider hints
- existing contract ids and expose ids
Forge does not send:
- raw sample rows
- full file contents
- secrets
- API keys
- passwords
How discovery affects generation
Discovery is combined with:
- CLI flags
- current-run answers
- optional project memory
- built-in defaults
That combined context shapes the contract draft and any follow-up prompts.