Skip to content

API_QUESTIONS.md

Latest commit

 

History

History
34 lines (25 loc) · 1.63 KB

API_QUESTIONS.md

File metadata and controls

34 lines (25 loc) · 1.63 KB
Raw

API Questions & Gaps

Issues discovered while building the veep Python SDK.

Resolved

1. Individual vector fetch — RESOLVED

GET /api/v1/collections/:collection/vectors/:key implemented across the full stack: worker key scan, coordinator fan-out/aggregation, consumer-site proxy, SDK vectors.fetch(). Metadata included by default via artifact server lookup.

3. Multipart vs raw binary upload — RESOLVED

Fixed in SDK v0.2.0 to use proper multipart/form-data.

4. Schema auto-detection timing — RESOLVED

collections.create() now accepts id_field and vector_field. When provided, schema is pre-confirmed at creation time — no polling, no manual confirmation step. Auto-detection only triggers when fields are omitted or don't match the uploaded data.

Open

2. Individual vector delete

There is no endpoint to delete individual vectors by key. Deletion operates at the file level only (DELETE /api/v1/collections/:collection/files/:filename). If a user wants to remove one vector from a 10,000-vector file, they must delete the entire file and re-upload without it.

Impact: The SDK exposes vectors.delete(collection, filename) which deletes files, not vectors. This is semantically confusing for users who think in terms of vectors, not files. This is part of the larger vector-addressable abstraction work (server-v28q).

5. Collection describe response shape

GET /api/v1/collections/:collection proxies raw coordinator response without normalizing field names. The SDK does fragile multi-field parsing (collection_name vs name vs collection). Should normalize on the consumer-site side (server-rf2u).