dazed/DazedTL
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Use array schema holy fucking god help me

Browse source
This commit is contained in:
dazedanon 2026-03-16 03:02:30 -05:00
parent 18bdf46f6b
commit 47ac9ac366
2 changed files with 35 additions and 16 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
prompt.txt
View file

@ -2,7 +2,7 @@ You are an expert Eroge game translator and localizer who translates Japanese te
You will be translating erotic and sexual content. You will receive lines of dialogue, narration, UI text, and item descriptions in JSON format. Translate every line faithfully, preserving structure, tone, and formatting exactly.
Test18
Test19
## Core Rules

49
util/translation.py
View file

@ -772,13 +772,28 @@ def createTranslationSchema(numLines):
return schema
def translateText(system, user, history, penalty, formatType, model, numLines=None, vocab_text="", batchSize=None):
# Static array-based schema for Claude's output_config. Because Anthropic
# includes output_config in the prompt-cache key, every call must send the
# exact same schema or the cache is busted. An array of strings naturally
# accommodates any batch size without changing.
_CLAUDE_TRANSLATION_SCHEMA = {
"type": "object",
"properties": {
"translations": {
"type": "array",
"items": {"type": "string"},
}
},
"required": ["translations"],
"additionalProperties": False,
}
def translateText(system, user, history, penalty, formatType, model, numLines=None, vocab_text=""):
"""Send translation request to the selected API.
system: Static system prompt (prompt.txt). Cached by Claude.
vocab_text: Per-batch vocabulary (dynamic, never cached to avoid cache busting).
batchSize: Max batch size — used for a fixed-size JSON schema so
output_config stays identical across Claude calls (cache stability).
"""
# Ensure system content is not empty
if not system or not str(system).strip():
@ -985,16 +1000,15 @@ def translateText(system, user, history, penalty, formatType, model, numLines=No
system=ant_system,
messages=native_msgs,
)
# Use a FIXED-SIZE schema (batchSize) so output_config is
# identical across all Claude calls — Anthropic includes it in
# the prompt-cache key, so varying numLines would bust the cache.
# The caller trims extra lines from the response after extraction.
# Use a STATIC array-based schema so output_config is identical
# across all Claude calls (Anthropic includes it in the prompt
# cache key). The caller sends LineN input; Claude returns
# {"translations": ["...", ...]} which extractTranslation handles.
if formatType == "json" and numLines is not None:
schema_lines = batchSize if batchSize else numLines
ant_kwargs["output_config"] = {
"format": {
"type": "json_schema",
"schema": createTranslationSchema(schema_lines),
"schema": _CLAUDE_TRANSLATION_SCHEMA,
}
}
@ -1224,6 +1238,11 @@ def extractTranslation(translatedTextList, isList, pbar=None):
try:
lineDict = json.loads(s)
# Handle array-based schema: {"translations": ["...", ...]}
if isinstance(lineDict, dict) and "translations" in lineDict and isinstance(lineDict["translations"], list):
stringList = [str(v) for v in lineDict["translations"]]
return stringList if isList else (stringList[0] if stringList else None)
# Build list in numeric order if keys are LineN
numeric_keys = []
for k in lineDict.keys():
@ -1598,7 +1617,7 @@ def translateAI(text, history, config, filename=None, pbar=None, lock=None, mism
# Translate
try:
response = translateText(static_system, current_user, history, 0.05, formatType, config.model, numLines, vocab_text=vocab_text, batchSize=config.batchSize)
response = translateText(static_system, current_user, history, 0.05, formatType, config.model, numLines, vocab_text=vocab_text)
except Exception as api_err:
err_msg = f"[API_ERROR] {api_err}"
# Print to stdout so the GUI captures it immediately
@ -1703,11 +1722,6 @@ def translateAI(text, history, config, filename=None, pbar=None, lock=None, mism
if isinstance(tItem, list):
extracted = extractTranslation(cleaned_text, True, pbar)
# Trim extra lines from fixed-size schema response
# (Claude fills all batchSize slots; we only need numLines).
if extracted and len(extracted) > len(clean_tItem):
extracted = extracted[:len(clean_tItem)]
# Check 1: Mismatch in length -> still a hard failure
if extracted is None or len(clean_tItem) != len(extracted):
is_valid = False
@ -1816,6 +1830,9 @@ def translateAI(text, history, config, filename=None, pbar=None, lock=None, mism
formatted_output = last_raw_translation
try:
parsed_json = json.loads(last_raw_translation)
# Normalize array-based output to LineN format for log readability
if isinstance(parsed_json, dict) and "translations" in parsed_json and isinstance(parsed_json["translations"], list):
parsed_json = {f"Line{i+1}": v for i, v in enumerate(parsed_json["translations"])}
formatted_output = json.dumps(parsed_json, indent=4, ensure_ascii=False)
except (json.JSONDecodeError, ValueError):
pass
@ -1858,6 +1875,8 @@ def translateAI(text, history, config, filename=None, pbar=None, lock=None, mism
formatted_mismatch_output = last_raw_translation
try:
parsed_json = json.loads(last_raw_translation)
if isinstance(parsed_json, dict) and "translations" in parsed_json and isinstance(parsed_json["translations"], list):
parsed_json = {f"Line{i+1}": v for i, v in enumerate(parsed_json["translations"])}
formatted_mismatch_output = json.dumps(parsed_json, indent=4, ensure_ascii=False)
except (json.JSONDecodeError, ValueError):
pass