Skip to content

Fix empty-output on hard batches: raise max_completion_tokens to 16384 (tunable) #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
msitarzewski merged 2 commits into from
Aug 26, 2026
Merged

Fix empty-output on hard batches: raise max_completion_tokens to 16384 (tunable) #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0703718
Fix empty output on hard batches: raise max_completion_tokens (16384,…
msitarzewski Aug 26, 2026
cf45b70
test: assert the 16384 max_completion_tokens default
msitarzewski Aug 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions fintick/aggregate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,10 @@
LLM_API_KEY = os.environ.get("FINTICK_LLM_API_KEY", "ollama")
LLM_MODEL = os.environ.get("FINTICK_LLM_MODEL", "qwen3.8:27b")
DEFAULT_MODEL = LLM_MODEL # backward-compat alias
# Reasoning models (GPT-5.x) spend this budget on hidden reasoning BEFORE emitting output, so it
# must leave room for both. 4096 gets fully consumed by reasoning on hard batches -> empty content;
# 16384 leaves ample headroom (observed ~5-7k total). Tunable per-model via the env var.
LLM_MAX_TOKENS = int(os.environ.get("FINTICK_LLM_MAX_TOKENS", "16384"))
WINDOW = timedelta(hours=6)
MAX_POSTS = 200
DEFAULT_BATCH = 50
Expand Down Expand Up @@ -416,8 +420,8 @@ def call_inference(
"model": model,
"stream": False,
# GPT-5.x requires max_completion_tokens (not max_tokens) and rejects any non-default
# temperature; response_format json_object constrains output regardless.
"max_completion_tokens": 4096,
# temperature; it also spends this budget on hidden reasoning first, so keep headroom.
"max_completion_tokens": LLM_MAX_TOKENS,
"response_format": {"type": "json_object"},
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
Expand Down
2 changes: 1 addition & 1 deletion tests/test_aggregate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,7 @@ def test_posts_openai_compatible_forced_json_request(self, urlopen: mock.Mock) -
self.assertEqual(payload["model"], "m-1")
self.assertIs(payload["stream"], False)
self.assertEqual(payload["response_format"], {"type": "json_object"})
self.assertLessEqual(payload["max_completion_tokens"], 4096)
self.assertEqual(payload["max_completion_tokens"], 16384)

@mock.patch("fintick.aggregate.urllib.request.urlopen")
def test_empty_content_raises(self, urlopen: mock.Mock) -> None:
Expand Down