/v1/vector_stores indexes UTF-8-decodable text only. A PDF, DOCX or PPTX attached to a
store settles as status=failed with last_error.code=unsupported_file, where the
upstream API indexes .pdf, .doc, .docx and .pptx alongside the code and text
formats.
PDF is the dominant real-world corpus for retrieval, so the limitation is the difference
between a store a customer can use and one they cannot. It is currently documented rather
than silent, which is the correct interim state.
Feature: extract document text before chunking, for at least PDF and DOCX. The repository
already ships a Docling-based ingestion sample, and the agentic suite already exercises
Docling against the gateway, so the extraction path has a proven precedent to follow.
Decide explicitly whether extraction runs in-process (a runtime dependency in both
container images, per stdapi-container-image) or via an AWS service, and record the
reason — the choice binds the image size and the cold-start budget.
/v1/vector_storesindexes UTF-8-decodable text only. A PDF, DOCX or PPTX attached to astore settles as
status=failedwithlast_error.code=unsupported_file, where theupstream API indexes
.pdf,.doc,.docxand.pptxalongside the code and textformats.
PDF is the dominant real-world corpus for retrieval, so the limitation is the difference
between a store a customer can use and one they cannot. It is currently documented rather
than silent, which is the correct interim state.
Feature: extract document text before chunking, for at least PDF and DOCX. The repository
already ships a Docling-based ingestion sample, and the agentic suite already exercises
Docling against the gateway, so the extraction path has a proven precedent to follow.
Decide explicitly whether extraction runs in-process (a runtime dependency in both
container images, per
stdapi-container-image) or via an AWS service, and record thereason — the choice binds the image size and the cold-start budget.