Formats, structured evidence, and export#

Readers normalize heterogeneous sources into CorpusDocument rather than making exporters understand every original file format.

Important provenance fields include the source path/URI, chunk identity, section/type metadata and reader-specific metadata such as timecodes or archive member paths.

Archives#

Archive extraction is bounded by file-count and total-size limits. Extracted members are routed through the same builder reader seam as normal files. A format-specific reader may be tried first for archive extensions that have a dedicated meaning; generic extraction is a deliberate fallback, not a silent replacement.

Export and adapters#

Use the export/adapters layer after canonical document creation rather than re-parsing source files. Existing adapters cover downstream formats such as LangChain/LangGraph/MCP and data exports.

When adding a new reader or exporter, keep source decoding and destination serialization separate so a new output format cannot alter ingestion semantics.