Vito Sansevero
a6d634509b
The joined-text refusal also refused the #170 parity files whose PromptManager prepend is a link the browser parser does not follow: its answer is then the node's literal text, which is what the node stores. Refuse only an answer that is not such a node's own text.
459 lines
16 KiB
Python
459 lines
16 KiB
Python
"""The per-file prompt decision shared by the output scan and the monitor (#247).
|
|
|
|
``decide_file_prompt`` is the ladder: which prompt one output file names,
|
|
from its metadata alone plus, when the graph cannot decide, the save node
|
|
recorded history says wrote it. It does no file or database I/O; the caller
|
|
reads the metadata, the file size and history (through ``lookup``) and
|
|
stores the ``Decision``.
|
|
|
|
The default resolver profile decides first, as ingestion always has. Only
|
|
for a file it leaves ambiguous or unresolved does the ladder ask the
|
|
inspector's resolve (``inspection_resolution``: the inspection profile, then
|
|
the recorded saver anchor; never the file-name saver anchor, whose files go
|
|
to Needs review by the #390 ruling), then the #170 browser parser (the legacy floor).
|
|
The policies differ in one place: for a file the inspector calls ambiguous,
|
|
the scan keeps the legacy floor and the monitor stores nothing. Junk text
|
|
(``junk_text``) is never stored, whichever step found it.
|
|
"""
|
|
|
|
import json
|
|
import re
|
|
from dataclasses import dataclass, field
|
|
from enum import Enum
|
|
from typing import Any, Callable, Mapping, Optional, Union
|
|
|
|
from .generation_params import (
|
|
PLACEHOLDERS,
|
|
GenerationParamsError,
|
|
extract_generation_params,
|
|
parse_page_json,
|
|
)
|
|
from .prompt_inspector import inspection_resolution
|
|
from .prompt_pair_resolver import _numeric, resolve_prompt_pair, sampler_less_prompt
|
|
from .review_capture import reviewable
|
|
|
|
try:
|
|
from ..database.pair_ingestion import blank_positive
|
|
except ImportError:
|
|
from database.pair_ingestion import blank_positive
|
|
|
|
|
|
class Policy(Enum):
|
|
SCAN = "scan"
|
|
MONITOR = "monitor"
|
|
|
|
|
|
PAIR = "pair"
|
|
POSITIVE = "positive"
|
|
HISTORY = "history"
|
|
NEVER = "never"
|
|
KINDS = (PAIR, POSITIVE, HISTORY, NEVER)
|
|
|
|
_UNDECIDED = ("ambiguous", "unresolved")
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class Decision:
|
|
"""What to store for one file (see the design's "The ladder").
|
|
|
|
``kind``: one of ``KINDS``. ``text``: the storage text of the positive.
|
|
``canonical``: the parsed metadata stored with the link. ``resolution``:
|
|
the resolution that decided a ``pair``, else the default profile's
|
|
result. ``source``: which step named the text (log lines and the scan's
|
|
outcome detail). ``review``: ``ambiguous`` or ``untraced`` when the
|
|
default profile was ambiguous (the scan's Needs review lists).
|
|
``try_recorded_run``: the scan tries the recorded run's stored prompt
|
|
when nothing else links. ``unlinked_status``: the scan status when
|
|
nothing links (None: no prompt metadata).
|
|
"""
|
|
|
|
kind: str
|
|
text: Optional[str] = None
|
|
canonical: dict = field(default_factory=dict)
|
|
resolution: Optional[dict] = None
|
|
source: str = "none"
|
|
review: Optional[str] = None
|
|
try_recorded_run: bool = False
|
|
unlinked_status: Optional[str] = None
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class _File:
|
|
"""The ladder's inputs beside the metadata."""
|
|
|
|
metadata: Mapping
|
|
canonical: dict
|
|
resolution: dict
|
|
review: Optional[str]
|
|
file_name: Optional[str]
|
|
in_output: bool
|
|
lookup: Optional[Callable[[], Any]]
|
|
policy: Policy
|
|
|
|
|
|
def decide_file_prompt(
|
|
metadata: Any,
|
|
*,
|
|
file_name: Optional[str],
|
|
file_size: Union[None, int, Callable[[], Optional[int]]],
|
|
in_output: bool,
|
|
lookup: Optional[Callable[[], Any]],
|
|
policy: Policy,
|
|
) -> Decision:
|
|
"""The ladder for one file's raw ``metadata``.
|
|
|
|
``prompt``/``workflow`` may be JSON strings (the scan) or parsed values
|
|
(the monitor). ``lookup`` returns the file's ``RecordedOutput`` or None;
|
|
it is None when history must not be asked yet. It is called at most
|
|
once, and only for a graph the default profile leaves undecided.
|
|
``file_size`` may be a callable: it is read only for a file with prompt
|
|
metadata whose ``file_info`` has no size.
|
|
"""
|
|
if not metadata:
|
|
return Decision(NEVER, try_recorded_run=True)
|
|
canonical = _parsed(metadata)
|
|
if not canonical.get("prompt") and not canonical.get("parameters"):
|
|
return Decision(NEVER, canonical=canonical, try_recorded_run=True)
|
|
_with_size(canonical, file_size)
|
|
resolution = resolve_prompt_pair(canonical)
|
|
if blank_positive(resolution):
|
|
# Only prepend/append text: no stored prompt can own it (#342)
|
|
resolution = {**resolution, "state": "unresolved"}
|
|
found = _File(
|
|
metadata,
|
|
canonical,
|
|
resolution,
|
|
_review(resolution),
|
|
file_name,
|
|
in_output,
|
|
lookup,
|
|
policy,
|
|
)
|
|
return _gated(_decide(found), found)
|
|
|
|
|
|
def _decide(found: _File) -> Decision:
|
|
state = found.resolution["state"]
|
|
graph = found.canonical.get("prompt")
|
|
if state == "resolved":
|
|
positive = found.resolution["positive"]
|
|
text = positive.get("storage_text", positive["text"])
|
|
return _decision(found, PAIR, text, "resolver")
|
|
if graph and state in _UNDECIDED:
|
|
return _undecided_graph(found)
|
|
if state == "ambiguous":
|
|
return _never(found, _unlinked(found))
|
|
if state == "unresolved" and not isinstance(found.canonical.get("parameters"), str):
|
|
return _never(found, "unresolved")
|
|
if graph:
|
|
return _sampler_less(found, graph)
|
|
text = _a1111_positive(found.metadata)
|
|
if text is None:
|
|
return _never(found, "found")
|
|
return _decision(found, POSITIVE, text, "a1111")
|
|
|
|
|
|
def _undecided_graph(found: _File) -> Decision:
|
|
recorded = found.lookup() if found.lookup is not None else None
|
|
saver = recorded.saver_id if recorded is not None else None
|
|
inspected = inspection_resolution(
|
|
found.metadata, found.file_name, saver, name_anchor=False
|
|
)
|
|
text = _inspected_text(inspected.resolution)
|
|
if text is not None:
|
|
source = "recorded-run" if inspected.by_record else "inspector"
|
|
return _decision(found, POSITIVE, text, source)
|
|
if found.policy is Policy.MONITOR:
|
|
if found.lookup is None and found.in_output:
|
|
return _decision(found, HISTORY, None, "none")
|
|
if not inspected.malformed and inspected.resolution["state"] == "ambiguous":
|
|
return _never(found, _unlinked(found))
|
|
text = graph_prompt_text(found.metadata)
|
|
if text is not None and _joined_text(found.canonical["prompt"], text):
|
|
# The legacy parser joins prepend + text + append: the text a
|
|
# PromptManager node stores cannot be told from it (#343).
|
|
return _never(found, _unlinked(found), try_recorded_run=True)
|
|
if text is not None:
|
|
return _decision(found, POSITIVE, text, "legacy-parser")
|
|
return _never(found, _unlinked(found), try_recorded_run=True)
|
|
|
|
|
|
def _sampler_less(found: _File, graph: Any) -> Decision:
|
|
# No node takes a positive or negative input (a guider or wrapper graph):
|
|
# the one wired text, never a guess (#252).
|
|
wired = sampler_less_prompt(graph)
|
|
if wired["text"] is not None:
|
|
return _decision(found, POSITIVE, wired["text"], "sampler-less")
|
|
if isinstance(found.canonical.get("parameters"), str):
|
|
# A metadata saver also wrote A1111 text: its positive prompt.
|
|
text = _a1111_positive(found.metadata)
|
|
if text is not None:
|
|
return _decision(found, POSITIVE, text, "a1111")
|
|
status = "unresolved" if wired["state"] == "ambiguous" else "found"
|
|
return _never(found, status, try_recorded_run=True)
|
|
|
|
|
|
_UNDECIDED_SOURCES = ("inspector", "recorded-run", "legacy-parser")
|
|
|
|
|
|
def _joined_text(graph: Any, text: str) -> bool:
|
|
"""Whether ``text`` may hold a PromptManager node's prepend or append.
|
|
|
|
True when a PromptManager-family node has a prepend or append part and
|
|
``text`` is not the literal ``text`` of one such node (the legacy
|
|
parser's answer when it does not follow a linked part).
|
|
"""
|
|
own = _prompt_manager_parts(graph)
|
|
return own is not None and text.strip() not in own
|
|
|
|
|
|
def _prompt_manager_parts(graph: Any) -> Optional[set]:
|
|
"""The literal texts of PromptManager-family nodes with a prepend or
|
|
append part, or None when no such node exists."""
|
|
if not isinstance(graph, dict):
|
|
return None
|
|
found = None
|
|
for node in graph.values():
|
|
if not isinstance(node, dict):
|
|
continue
|
|
name, inputs = node.get("class_type"), node.get("inputs")
|
|
if not isinstance(name, str) or "promptmanager" not in name.lower():
|
|
continue
|
|
if not isinstance(inputs, dict) or not _has_part(inputs):
|
|
continue
|
|
found = found if found is not None else set()
|
|
if isinstance(inputs.get("text"), str):
|
|
found.add(inputs["text"].strip())
|
|
return found
|
|
|
|
|
|
def _has_part(inputs: dict) -> bool:
|
|
for key in ("prepend_text", "append_text"):
|
|
part = inputs.get(key)
|
|
if part is not None and not (isinstance(part, str) and not part.strip()):
|
|
return True
|
|
return False
|
|
|
|
|
|
def _gated(decision: Decision, found: _File) -> Decision:
|
|
"""Junk text is never stored (step 13).
|
|
|
|
A junk answer of the undecided-graph step keeps that step's unlinked
|
|
status, so an ambiguous file is still a Needs review case; any other
|
|
step reports ``found``.
|
|
"""
|
|
if decision.kind not in (PAIR, POSITIVE):
|
|
return decision
|
|
junk = junk_text(decision.text)
|
|
if junk is None:
|
|
return decision
|
|
return Decision(
|
|
NEVER,
|
|
canonical=decision.canonical,
|
|
resolution=decision.resolution,
|
|
source=f"junk:{junk}",
|
|
review=decision.review,
|
|
try_recorded_run=True,
|
|
unlinked_status=(
|
|
_unlinked(found) if decision.source in _UNDECIDED_SOURCES else "found"
|
|
),
|
|
)
|
|
|
|
|
|
def _decision(found: _File, kind: str, text: Optional[str], source: str) -> Decision:
|
|
return Decision(
|
|
kind,
|
|
text=text,
|
|
canonical=found.canonical,
|
|
resolution=found.resolution,
|
|
source=source,
|
|
review=found.review,
|
|
)
|
|
|
|
|
|
def _never(found: _File, status: str, *, try_recorded_run: bool = False) -> Decision:
|
|
return Decision(
|
|
NEVER,
|
|
canonical=found.canonical,
|
|
resolution=found.resolution,
|
|
review=found.review,
|
|
try_recorded_run=try_recorded_run,
|
|
unlinked_status=status,
|
|
)
|
|
|
|
|
|
def _review(resolution: dict) -> Optional[str]:
|
|
if resolution["state"] != "ambiguous":
|
|
return None
|
|
return "ambiguous" if reviewable(resolution) else "untraced"
|
|
|
|
|
|
def _unlinked(found: _File) -> str:
|
|
"""Ambiguous without a choice is reported unresolved."""
|
|
if found.review == "untraced":
|
|
return "unresolved"
|
|
return found.resolution["state"]
|
|
|
|
|
|
def _inspected_text(resolution: dict) -> Optional[str]:
|
|
"""The storage text of a positive the inspector shows, else None."""
|
|
if resolution.get("state") in _UNDECIDED:
|
|
return None
|
|
positive = resolution.get("positive")
|
|
if not isinstance(positive, dict):
|
|
return None
|
|
text = positive.get("storage_text", positive.get("text"))
|
|
return text if isinstance(text, str) and text.strip() else None
|
|
|
|
|
|
def _parsed(metadata: Mapping) -> dict:
|
|
canonical = dict(metadata)
|
|
for key in ("prompt", "workflow"):
|
|
if isinstance(canonical.get(key), str):
|
|
try:
|
|
canonical[key] = json.loads(canonical[key])
|
|
except (ValueError, TypeError):
|
|
# Preserve malformed metadata for conservative resolver rejection.
|
|
pass
|
|
return canonical
|
|
|
|
|
|
def _with_size(canonical: dict, file_size: Any) -> None:
|
|
"""Record the file's size in ``file_info`` unless the metadata carries one.
|
|
|
|
Stored as ``generated_images.file_size``; the "new" scan scope skips a
|
|
known file only while its size is unchanged.
|
|
"""
|
|
info = canonical.get("file_info")
|
|
info = dict(info) if isinstance(info, dict) else {}
|
|
if info.get("size") is None:
|
|
size = file_size() if callable(file_size) else file_size
|
|
if size is not None:
|
|
info["size"] = size
|
|
if info:
|
|
canonical["file_info"] = info
|
|
|
|
|
|
def _a1111_positive(metadata: Mapping) -> Optional[str]:
|
|
text = parse_comfyui_prompt(metadata)["positive_prompt"]
|
|
return text.strip() if text and text.strip() else None
|
|
|
|
|
|
_LORA_TAG = re.compile(r"<lora:[^>]*>")
|
|
_LORA_SEPARATORS = " \t\r\n\f\v,;|"
|
|
_MODEL_FILE = re.compile(r"\S*\.(?:safetensors|ckpt|pt|pth|bin|gguf|sft)", re.I)
|
|
|
|
|
|
def junk_text(text: Any) -> Optional[str]:
|
|
"""The #252 junk class of ``text``, or None when it may be a prompt."""
|
|
stripped = text.strip() if isinstance(text, str) else ""
|
|
if not stripped:
|
|
return "blank"
|
|
if _numeric(stripped):
|
|
return "number"
|
|
if not _LORA_TAG.sub("", stripped).strip(_LORA_SEPARATORS):
|
|
return "lora-only"
|
|
if _MODEL_FILE.fullmatch(stripped):
|
|
return "model-file"
|
|
return None
|
|
|
|
|
|
def graph_prompt_text(metadata: Mapping) -> Optional[str]:
|
|
"""The positive prompt the browser parser traces from the graphs, or None.
|
|
|
|
Records parse as the legacy pages parsed them (NaN reads as null).
|
|
"""
|
|
records = {}
|
|
for key in ("prompt", "workflow"):
|
|
value = metadata.get(key)
|
|
records[key] = parse_page_json(value) if isinstance(value, str) else value
|
|
try:
|
|
positive = extract_generation_params(records)["positivePrompt"]
|
|
except GenerationParamsError:
|
|
return None
|
|
if not isinstance(positive, str) or positive == PLACEHOLDERS["positivePrompt"]:
|
|
return None
|
|
return positive.strip() or None
|
|
|
|
|
|
_COMMON_FIELDS = (
|
|
"positive",
|
|
"negative",
|
|
"steps",
|
|
"cfg",
|
|
"sampler",
|
|
"scheduler",
|
|
"seed",
|
|
)
|
|
|
|
|
|
def _json_or_raw(value: Any) -> Any:
|
|
if not isinstance(value, str):
|
|
return value
|
|
try:
|
|
return json.loads(value)
|
|
except json.JSONDecodeError:
|
|
return value
|
|
|
|
|
|
def parse_comfyui_prompt(metadata: Mapping) -> dict:
|
|
"""Parse ComfyUI and A1111 prompt data from metadata.
|
|
|
|
Accepts the file's strings or parsed values: a parsed ``prompt`` or
|
|
``workflow`` is kept as it is and a non-string ``parameters`` is ignored.
|
|
"""
|
|
result = {
|
|
"prompt": None,
|
|
"workflow": None,
|
|
"parameters": {},
|
|
"positive_prompt": None,
|
|
"negative_prompt": None,
|
|
}
|
|
params = metadata.get("parameters")
|
|
if isinstance(params, str):
|
|
_parse_a1111(params, result)
|
|
if result["positive_prompt"] is None:
|
|
for key in ("prompt", "workflow"):
|
|
if key in metadata:
|
|
result[key] = _json_or_raw(metadata[key])
|
|
for name in _COMMON_FIELDS:
|
|
if name in metadata:
|
|
result["parameters"][name] = _json_or_raw(metadata[name])
|
|
return result
|
|
|
|
|
|
_SETTINGS_FIELDS = ("Steps", "Sampler", "Schedule type", "CFG scale", "Seed")
|
|
|
|
|
|
def _parse_a1111(params: str, result: dict) -> None:
|
|
"""A1111/Forge ``parameters``: a possibly multi-line positive, then a
|
|
"Negative prompt:" section, then a trailing "Steps: ..." settings line.
|
|
Taking the first line as the positive would truncate a multi-line prompt,
|
|
so the text is split on the settings line instead.
|
|
"""
|
|
body = params
|
|
settings = ""
|
|
found = re.search(r"\nSteps:\s", body)
|
|
if found:
|
|
settings = body[found.start() + 1 :].strip()
|
|
body = body[: found.start()]
|
|
negative = None
|
|
found = re.search(r"(?im)^\s*negative prompt:\s*", body)
|
|
if found:
|
|
negative = body[found.end() :].strip()
|
|
body = body[: found.start()]
|
|
elif body.lower().lstrip().startswith("negative prompt:"):
|
|
negative = body.lower().lstrip()[len("negative prompt:") :].strip()
|
|
body = ""
|
|
result["positive_prompt"] = body.strip() or None
|
|
if negative:
|
|
result["negative_prompt"] = negative
|
|
result["parameters"]["parameters"] = params
|
|
for name in _SETTINGS_FIELDS if settings else ():
|
|
found = re.search(
|
|
r"(?i)(?:^|,\s*)" + re.escape(name) + r":\s*([^,]+)", settings
|
|
)
|
|
if found:
|
|
key = name.lower().replace(" ", "_")
|
|
result["parameters"][key] = found.group(1).strip()
|