Add transitional learning detect with in-dialog mask review.
Deploy Development / deploy (push) Successful in 57s
Test Suite / pytest-backend (push) Failing after 2m57s
Test Suite / smoke-dev (push) Successful in 0s
Test Suite / frontend-build (push) Successful in 16s

Senses grow from confirmations instead of a word list, so a later local pipeline can decide homonyms on a short passage. Default stays semantic. Local detect waits longer, and the reverse-proxy timeout is documented so Dev does not 504 first.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
LarsandCursor committed 2026-09-09 08:34:43 +02:00
1 parent 59921fc5b5
commit fa1c11c33e
29 files changed
+1135 -45

No files matched your search

+12 -2
View File
@@ -31,7 +31,9 @@ from providers import ChatResult, ProviderError, complete_chat, detect_provider
DETECT_CHUNK_CHARS = 4000
DETECT_CHUNK_OVERLAP = 250
DETECT_TIMEOUT = 90.0
DETECT_TIMEOUT_LOCAL = 300.0
DETECT_MAX_TOKENS = 1024
DETECT_MAX_TOKENS_LOCAL = 256
JSON_BLOCK = re.compile(r"\{.*\}", re.DOTALL)
ALLOWED_ENTITY_FIELDS = frozenset({"start", "end", "text", "entity_type"})
ALLOWED_ROOT_FIELDS = frozenset({"entities"})
@@ -573,6 +575,13 @@ def _add_usage(stats: DetectionStats, usage: dict | None) -> None:
pass
def detect_chat_limits(config) -> tuple[float, int]:
"""Remote detect stays at 90s/1024. Local CPU hosts need a longer wait and a shorter cap."""
if getattr(config, "local", False):
return DETECT_TIMEOUT_LOCAL, DETECT_MAX_TOKENS_LOCAL
return DETECT_TIMEOUT, DETECT_MAX_TOKENS
def _llm_chunk(config, excerpt: str, *, schema_retry: bool = False) -> ChatResult:
prompt = resolve_template(
_detect_prompt()["template"],
@@ -580,11 +589,12 @@ def _llm_chunk(config, excerpt: str, *, schema_retry: bool = False) -> ChatResul
)
if schema_retry:
prompt = f"{prompt.rstrip()}\n\n{SCHEMA_RETRY_HINT}"
timeout, max_tokens = detect_chat_limits(config)
return complete_chat(
config,
[{"role": "user", "content": prompt}],
timeout=DETECT_TIMEOUT,
max_tokens=DETECT_MAX_TOKENS,
timeout=timeout,
max_tokens=max_tokens,
disable_context_compression=True,
)