Distill and infer from documents

Distill a directory

construct distill reads a directory of markdown, code, or text files and returns a context-budget-friendly summary focused on a specific question.

construct distill docs/
construct distill docs/ --query="what are the accepted ADR decisions?"

Without --query, it produces a general summary. With --query, chunk selection is focused on answering that question specifically: useful before a long coding session to pre-load only the relevant context.

Output formats

construct distill docs/ --format=summary       # default: paragraph summary
construct distill docs/ --format=decisions     # bullet list of decisions only
construct distill docs/ --format=full          # all chunks, ranked by relevance
construct distill docs/ --format=extract       # raw chunk text, no synthesis

Write output to a file

construct distill docs/ --query="storage layer" --out=.construct/distill-storage.md

The file can be passed as context to a fresh agent session or committed alongside a plan.

Control scan depth and file types

construct distill src/ --depth=2 --ext=ts,tsx

--depth limits directory recursion. --ext restricts to specific extensions.

---

Infer a document schema

construct infer sends a document (or a batch of documents) to the model and returns a structured field schema: field names, types, formats, examples, and confidence scores.

Single document

construct infer docs/specs/prd/prd-001-storage.md

Output is a JSON schema printed to stdout.

Multiple documents (unified schema)

construct infer docs/specs/prd/*.md --unified

Reconciles fields across all matched files. Fields that appear in fewer than 50% of documents are excluded by default.

Flag	Effect
--unified	Reconcile fields across multiple files into one schema
--max-chars=N	Max characters per document sent to the model (default: 40000)
--sample=N	Max documents to sample in unified mode (default: 10)
--threshold=0.5	Field inclusion threshold for unified mode (0.0–1.0)

Save the schema as an artifact

The MCP tool infer_document_schema accepts a save: true flag that writes the result to .construct/knowledge/reference/schemas/\<name>.schema.json. This is not available via the CLI flag yet: use the MCP surface directly when saving to the schema registry is needed.