mirror of
https://github.com/warmbly/warmbly.git
synced 2026-09-11 16:08:09 +00:00
Matthew Meszaros
2009780beb
Add OpenAI Batch API support for warmup content generation, including job metadata, polling, cancellation, and completed-batch ingestion. Share the generation prompt between sync and batch modes and humanize/lint generated threads before storing them in the warmup content library.
26 lines
1.4 KiB
SQL
26 lines
1.4 KiB
SQL
-- Warmup batch generation.
|
|
--
|
|
-- The offline warmup-content generator originally ran every model call
|
|
-- synchronously (one chat-completion per thread). The OpenAI Batch API is ~50%
|
|
-- cheaper and processes asynchronously (up to a 24h window), which is a much
|
|
-- better fit for bulk thread-bank refills where latency does not matter.
|
|
--
|
|
-- This migration extends warmup_generation_jobs so one job row can represent
|
|
-- either a synchronous run (mode='sync', the existing behaviour) or an async
|
|
-- batch run (mode='batch'). For batch runs we track the OpenAI batch ID, the
|
|
-- uploaded input file ID, the output file ID (populated when the batch
|
|
-- completes), the last-observed OpenAI batch status, and the requested
|
|
-- completion window. A background poller reconciles in-flight batch jobs against
|
|
-- the OpenAI batch status and ingests results when they finish.
|
|
|
|
ALTER TABLE ONLY public.warmup_generation_jobs
|
|
ADD COLUMN mode text NOT NULL DEFAULT 'sync',
|
|
ADD COLUMN batch_id text NOT NULL DEFAULT '',
|
|
ADD COLUMN batch_input_file_id text NOT NULL DEFAULT '',
|
|
ADD COLUMN batch_output_file_id text NOT NULL DEFAULT '',
|
|
ADD COLUMN batch_status text NOT NULL DEFAULT '',
|
|
ADD COLUMN completion_window text NOT NULL DEFAULT '24h';
|
|
|
|
-- The poller scans for in-flight batch jobs by mode + status, so index that.
|
|
CREATE INDEX idx_warmup_generation_jobs_batch ON public.warmup_generation_jobs USING btree (mode, status, batch_status);
|