mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
- A completion is the exact continuation of its prompt, so it keeps the whitespace its
first token carries (' Paris', not 'Paris'); only chat messages drop it, in the final
text and the first streamed delta. Clients concatenating prompt and completion got
'isParis'.
- /v1/completions logprobs come in the completions shape OpenAI clients read (tokens,
token_logprobs, top_logprobs, text_offset), for the full choice and each streamed
chunk, instead of the chat shape.
Co-authored-by: Hanzo Dev <[email protected]>