1038 lines
46 KiB
Python
1038 lines
46 KiB
Python
from __future__ import annotations
|
||
|
||
from typing import Any, Dict, List, Optional, Callable, Awaitable
|
||
from urllib.parse import urljoin
|
||
import json
|
||
import re
|
||
import asyncio
|
||
import time
|
||
from agentui.providers.http_client import build_client
|
||
from agentui.common.vendors import detect_vendor
|
||
from agentui.pipeline.templating import (
|
||
_OUT_MACRO_RE,
|
||
_VAR_MACRO_RE,
|
||
_PROMPT_MACRO_RE,
|
||
_OUT_SHORT_RE,
|
||
_BRACES_RE,
|
||
_split_path,
|
||
_get_by_path,
|
||
_stringify_for_template,
|
||
_deep_find_text,
|
||
_best_text_from_outputs,
|
||
render_template_simple,
|
||
)
|
||
|
||
|
||
# --- Templating helpers are imported from agentui.pipeline.templating ---
|
||
|
||
# moved to agentui.pipeline.templating
|
||
|
||
# moved to agentui.pipeline.templating
|
||
|
||
# moved to agentui.pipeline.templating
|
||
|
||
# moved to agentui.pipeline.templating
|
||
|
||
def _extract_out_node_id_from_ref(s: Any) -> Optional[str]:
|
||
"""
|
||
Извлекает node_id из строки с макросом [[OUT:nodeId(.path)*]].
|
||
Возвращает None, если макрос не найден.
|
||
"""
|
||
if not isinstance(s, str):
|
||
return None
|
||
m = _OUT_MACRO_RE.search(s)
|
||
if not m:
|
||
return None
|
||
body = m.group(1).strip()
|
||
node_id = body.split(".", 1)[0].strip()
|
||
return node_id or None
|
||
|
||
|
||
def _resolve_in_value(source: Any, context: Dict[str, Any], values: Dict[str, Dict[str, Any]]) -> Any:
|
||
"""
|
||
Разрешает входные связи/макросы в значение для inputs:
|
||
- Нестроковые значения возвращаются как есть
|
||
- "macro:path" → берёт значение из context по точечному пути
|
||
- "[[VAR:path]]" → берёт значение из context
|
||
- "[[OUT:nodeId(.path)*]]" → берёт из уже вычисленных выходов ноды
|
||
- "nodeId(.path)*" → ссылка на выходы ноды
|
||
- Если передан список ссылок — вернёт список разрешённых значений
|
||
- Иначе пытается взять из context по пути; если не найдено, оставляет исходную строку
|
||
"""
|
||
# Поддержка массивов ссылок (для multi-depends или будущих списковых входов)
|
||
if isinstance(source, list):
|
||
return [_resolve_in_value(s, context, values) for s in source]
|
||
if not isinstance(source, str):
|
||
return source
|
||
s = source.strip()
|
||
|
||
# macro:path
|
||
if s.lower().startswith("macro:"):
|
||
path = s.split(":", 1)[1].strip()
|
||
return _get_by_path(context, path)
|
||
|
||
# [[VAR: path]]
|
||
m = _VAR_MACRO_RE.fullmatch(s)
|
||
if m:
|
||
path = m.group(1).strip()
|
||
return _get_by_path(context, path)
|
||
|
||
# [[OUT: nodeId(.path)*]]
|
||
m = _OUT_MACRO_RE.fullmatch(s)
|
||
if m:
|
||
body = m.group(1).strip()
|
||
if "." in body:
|
||
node_id, rest = body.split(".", 1)
|
||
node_val = values.get(node_id)
|
||
return _get_by_path(node_val, rest)
|
||
node_val = values.get(body)
|
||
return node_val
|
||
|
||
# "nodeId(.path)*"
|
||
if "." in s:
|
||
node_id, rest = s.split(".", 1)
|
||
if node_id in values:
|
||
return _get_by_path(values.get(node_id), rest)
|
||
if s in values:
|
||
return values.get(s)
|
||
|
||
# fallback: from context by dotted path or raw string
|
||
ctx_val = _get_by_path(context, s)
|
||
return ctx_val if ctx_val is not None else source
|
||
|
||
|
||
# moved to agentui.pipeline.templating
|
||
|
||
|
||
class ExecutionError(Exception):
|
||
pass
|
||
|
||
|
||
class Node:
|
||
type_name: str = "Base"
|
||
|
||
def __init__(self, node_id: str, config: Optional[Dict[str, Any]] = None) -> None:
|
||
self.node_id = node_id
|
||
self.config = config or {}
|
||
|
||
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
|
||
"""Execute node with inputs and context. Return dict of outputs."""
|
||
raise NotImplementedError
|
||
|
||
# Регистрация поддерживаемых типов нод (минимальный набор)
|
||
NODE_REGISTRY: Dict[str, Any] = {}
|
||
|
||
|
||
class PipelineExecutor:
|
||
def __init__(self, pipeline: Dict[str, Any]) -> None:
|
||
self.pipeline = pipeline
|
||
self.nodes_by_id: Dict[str, Node] = {}
|
||
for n in pipeline.get("nodes", []):
|
||
node_cls = NODE_REGISTRY.get(n.get("type"))
|
||
if not node_cls:
|
||
raise ExecutionError(f"Unknown node type: {n.get('type')}")
|
||
self.nodes_by_id[n["id"]] = node_cls(n["id"], n.get("config", {}))
|
||
|
||
async def run(
|
||
self,
|
||
context: Dict[str, Any],
|
||
trace: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
|
||
) -> Dict[str, Any]:
|
||
"""
|
||
Исполнитель пайплайна с динамическим порядком на основе зависимостей графа.
|
||
Новый режим: волновое (level-by-level) исполнение с параллелизмом и барьером.
|
||
Все узлы «готовой волны» стартуют параллельно, ждём всех, затем открывается следующая волна.
|
||
Ограничение параллелизма берётся из pipeline.parallel_limit (по умолчанию 8).
|
||
Политика ошибок: fail-fast — при исключении любой задачи волны прерываем пайплайн.
|
||
"""
|
||
nodes: List[Dict[str, Any]] = list(self.pipeline.get("nodes", []))
|
||
id_set = set(self.nodes_by_id.keys())
|
||
|
||
# Собираем зависимости: node_id -> set(parent_ids), и обратные связи dependents
|
||
deps_map: Dict[str, set] = {n["id"]: set() for n in nodes}
|
||
dependents: Dict[str, set] = {n["id"]: set() for n in nodes}
|
||
|
||
for n in nodes:
|
||
nid = n["id"]
|
||
for _, source in (n.get("in") or {}).items():
|
||
# Разворачиваем массивы ссылок (multi-depends)
|
||
sources = source if isinstance(source, list) else [source]
|
||
for src in sources:
|
||
if not isinstance(src, str):
|
||
# Нестрочные значения и массивы констант — зависимостей нет
|
||
continue
|
||
if src.startswith("macro:"):
|
||
# Макросы берутся из контекста, без зависимостей
|
||
continue
|
||
# [[VAR:...]] — макрос из контекста, зависимостей нет
|
||
if re.fullmatch(r"\[\[\s*VAR\s*[:\s]\s*[^\]]+\s*\]\]", src.strip()):
|
||
continue
|
||
# [[OUT:nodeId(.key)*]] — зависимость от указанной ноды
|
||
out_ref_node = _extract_out_node_id_from_ref(src)
|
||
if out_ref_node and out_ref_node in id_set:
|
||
deps_map[nid].add(out_ref_node)
|
||
dependents[out_ref_node].add(nid)
|
||
continue
|
||
# Ссылки вида "node.outKey" или "node"
|
||
src_id = src.split(".", 1)[0] if "." in src else src
|
||
if src_id in id_set:
|
||
deps_map[nid].add(src_id)
|
||
dependents[src_id].add(nid)
|
||
|
||
# Входящие степени и первая волна
|
||
in_degree: Dict[str, int] = {nid: len(deps) for nid, deps in deps_map.items()}
|
||
ready: List[str] = [nid for nid, deg in in_degree.items() if deg == 0]
|
||
|
||
processed: List[str] = []
|
||
values: Dict[str, Dict[str, Any]] = {}
|
||
last_result: Dict[str, Any] = {}
|
||
node_def_by_id: Dict[str, Dict[str, Any]] = {n["id"]: n for n in nodes}
|
||
# Накопитель пользовательских переменных (SetVars) — доступен как context["vars"]
|
||
user_vars: Dict[str, Any] = {}
|
||
|
||
# Параметры параллелизма
|
||
try:
|
||
parallel_limit = int(self.pipeline.get("parallel_limit", 8))
|
||
except Exception:
|
||
parallel_limit = 8
|
||
if parallel_limit <= 0:
|
||
parallel_limit = 1
|
||
|
||
# Вспомогательная корутина исполнения одной ноды со снапшотом OUT
|
||
async def exec_one(node_id: str, values_snapshot: Dict[str, Any], wave_num: int) -> tuple[str, Dict[str, Any]]:
|
||
ndef = node_def_by_id.get(node_id)
|
||
if not ndef:
|
||
raise ExecutionError(f"Node definition not found: {node_id}")
|
||
node = self.nodes_by_id[node_id]
|
||
|
||
# Снимок контекста и OUT на момент старта волны
|
||
ctx = dict(context)
|
||
ctx["OUT"] = values_snapshot
|
||
# Пользовательские переменные (накопленные SetVars)
|
||
try:
|
||
ctx["vars"] = dict(user_vars)
|
||
except Exception:
|
||
ctx["vars"] = {}
|
||
|
||
# Разрешаем inputs для ноды
|
||
inputs: Dict[str, Any] = {}
|
||
for name, source in (ndef.get("in") or {}).items():
|
||
if isinstance(source, list):
|
||
inputs[name] = [_resolve_in_value(s, ctx, values_snapshot) for s in source]
|
||
else:
|
||
inputs[name] = _resolve_in_value(source, ctx, values_snapshot)
|
||
|
||
# Трассировка старта
|
||
if trace is not None:
|
||
try:
|
||
await trace({"event": "node_start", "node_id": ndef["id"], "wave": wave_num, "ts": int(time.time() * 1000)})
|
||
except Exception:
|
||
pass
|
||
|
||
started = time.perf_counter()
|
||
try:
|
||
out = await node.run(inputs, ctx)
|
||
except Exception as exc:
|
||
if trace is not None:
|
||
try:
|
||
await trace({
|
||
"event": "node_error",
|
||
"node_id": ndef["id"],
|
||
"wave": wave_num,
|
||
"ts": int(time.time() * 1000),
|
||
"error": str(exc),
|
||
})
|
||
except Exception:
|
||
pass
|
||
raise
|
||
else:
|
||
dur_ms = int((time.perf_counter() - started) * 1000)
|
||
if trace is not None:
|
||
try:
|
||
await trace({
|
||
"event": "node_done",
|
||
"node_id": ndef["id"],
|
||
"wave": wave_num,
|
||
"ts": int(time.time() * 1000),
|
||
"duration_ms": dur_ms,
|
||
})
|
||
except Exception:
|
||
pass
|
||
return node_id, out
|
||
|
||
# Волновое исполнение
|
||
wave_idx = 0
|
||
while ready:
|
||
wave_nodes = list(ready)
|
||
ready = [] # будет заполнено после завершения волны
|
||
wave_results: Dict[str, Dict[str, Any]] = {}
|
||
# Один общий снапшот OUT для всей волны (барьер — узлы волны не видят результаты друг друга)
|
||
values_snapshot = dict(values)
|
||
|
||
# Чанковый запуск с лимитом parallel_limit
|
||
for i in range(0, len(wave_nodes), parallel_limit):
|
||
chunk = wave_nodes[i : i + parallel_limit]
|
||
# fail-fast: при исключении любой задачи gather бросит и отменит остальные
|
||
results = await asyncio.gather(
|
||
*(exec_one(nid, values_snapshot, wave_idx) for nid in chunk),
|
||
return_exceptions=False,
|
||
)
|
||
# Коммитим результаты чанка в локальное хранилище волны
|
||
for nid, out in results:
|
||
wave_results[nid] = out
|
||
last_result = out # обновляем на каждом успешном результате
|
||
|
||
# После завершения волны — коммитим все её результаты в общие values
|
||
values.update(wave_results)
|
||
processed.extend(wave_nodes)
|
||
|
||
# Соберём пользовательские переменные из SetVars узлов волны
|
||
try:
|
||
for _nid, out in wave_results.items():
|
||
if isinstance(out, dict):
|
||
v = out.get("vars")
|
||
if isinstance(v, dict):
|
||
user_vars.update(v)
|
||
except Exception:
|
||
pass
|
||
|
||
# Обновляем входящие степени для зависимых и формируем следующую волну
|
||
for done in wave_nodes:
|
||
for child in dependents.get(done, ()):
|
||
in_degree[child] -= 1
|
||
next_ready = [nid for nid, deg in in_degree.items() if deg == 0 and nid not in processed and nid not in wave_nodes]
|
||
# Исключаем уже учтённые и добавляем только те, которые действительно готовы
|
||
ready = next_ready
|
||
wave_idx += 1
|
||
|
||
# Проверка на циклы/недостижимые ноды
|
||
if len(processed) != len(nodes):
|
||
remaining = [n["id"] for n in nodes if n["id"] not in processed]
|
||
raise ExecutionError(f"Cycle detected or unresolved dependencies among nodes: {remaining}")
|
||
|
||
return last_result
|
||
|
||
|
||
class SetVarsNode(Node):
|
||
type_name = "SetVars"
|
||
|
||
def _normalize(self) -> List[Dict[str, Any]]:
|
||
raw = self.config.get("variables") or []
|
||
if not isinstance(raw, list):
|
||
return []
|
||
norm: List[Dict[str, Any]] = []
|
||
for i, b in enumerate(raw):
|
||
if not isinstance(b, dict):
|
||
continue
|
||
name = str(b.get("name", "")).strip()
|
||
mode = str(b.get("mode", "string")).lower().strip()
|
||
value = b.get("value", "")
|
||
try:
|
||
order = int(b.get("order")) if b.get("order") is not None else i
|
||
except Exception:
|
||
order = i
|
||
norm.append({
|
||
"id": b.get("id") or f"v{i}",
|
||
"name": name,
|
||
"mode": "expr" if mode == "expr" else "string",
|
||
"value": value,
|
||
"order": order,
|
||
})
|
||
return norm
|
||
|
||
def _safe_eval_expr(self, expr: str) -> Any:
|
||
"""
|
||
Безопасная оценка выражений для SetVars.
|
||
|
||
Поддержка:
|
||
- Литералы: числа/строки/bool/None, списки, кортежи, словари
|
||
- JSON‑литералы: true/false/null, объекты и массивы (парсятся как Python True/False/None, dict/list)
|
||
- Арифметика: + - * / // %, унарные +-
|
||
- Логика: and/or, сравнения (== != < <= > >=, цепочки)
|
||
- Безопасные функции: rand(), randint(a,b), choice(list)
|
||
|
||
Запрещено: имя/атрибуты/индексация/условные/импорты/прочие вызовы функций.
|
||
"""
|
||
import ast
|
||
import operator as op
|
||
import random
|
||
|
||
# 0) Попытаться распознать чистый JSON‑литерал (включая true/false/null, объекты/массивы/числа/строки).
|
||
# Это не вмешивается в математику: для выражений вида "1+2" json.loads бросит исключение и мы пойдём в AST.
|
||
try:
|
||
s = str(expr).strip()
|
||
return json.loads(s)
|
||
except Exception:
|
||
pass
|
||
|
||
allowed_bin = {
|
||
ast.Add: op.add, ast.Sub: op.sub, ast.Mult: op.mul, ast.Div: op.truediv,
|
||
ast.FloorDiv: op.floordiv, ast.Mod: op.mod,
|
||
}
|
||
allowed_unary = {ast.UAdd: lambda x: +x, ast.USub: lambda x: -x}
|
||
allowed_cmp = {
|
||
ast.Eq: op.eq, ast.NotEq: op.ne, ast.Lt: op.lt, ast.LtE: op.le, ast.Gt: op.gt, ast.GtE: op.ge,
|
||
}
|
||
|
||
def eval_node(node: ast.AST) -> Any:
|
||
if isinstance(node, ast.Expression):
|
||
return eval_node(node.body)
|
||
if isinstance(node, ast.Constant):
|
||
return node.value
|
||
if isinstance(node, ast.Tuple):
|
||
return tuple(eval_node(e) for e in node.elts)
|
||
if isinstance(node, ast.List):
|
||
return [eval_node(e) for e in node.elts]
|
||
if isinstance(node, ast.Dict):
|
||
return {eval_node(k): eval_node(v) for k, v in zip(node.keys, node.values)}
|
||
if isinstance(node, ast.UnaryOp) and type(node.op) in allowed_unary:
|
||
return allowed_unary[type(node.op)](eval_node(node.operand))
|
||
if isinstance(node, ast.BinOp) and type(node.op) in allowed_bin:
|
||
return allowed_bin[type(node.op)](eval_node(node.left), eval_node(node.right))
|
||
if isinstance(node, ast.BoolOp):
|
||
vals = [eval_node(v) for v in node.values]
|
||
if isinstance(node.op, ast.And):
|
||
res = True
|
||
for v in vals:
|
||
res = res and bool(v)
|
||
return res
|
||
if isinstance(node.op, ast.Or):
|
||
res = False
|
||
for v in vals:
|
||
res = res or bool(v)
|
||
return res
|
||
if isinstance(node, ast.Compare):
|
||
left = eval_node(node.left)
|
||
for opnode, comparator in zip(node.ops, node.comparators):
|
||
if type(opnode) not in allowed_cmp:
|
||
raise ExecutionError("Unsupported comparison operator")
|
||
right = eval_node(comparator)
|
||
if not allowed_cmp[type(opnode)](left, right):
|
||
return False
|
||
left = right
|
||
return True
|
||
# Разрешённые вызовы: rand(), randint(a,b), choice(list)
|
||
if isinstance(node, ast.Call):
|
||
# Никаких kwargs, *args
|
||
if node.keywords or isinstance(getattr(node, "starargs", None), ast.AST) or isinstance(getattr(node, "kwargs", None), ast.AST):
|
||
raise ExecutionError("Call with kwargs/starargs is not allowed")
|
||
fn = node.func
|
||
if not isinstance(fn, ast.Name):
|
||
raise ExecutionError("Only simple function calls are allowed")
|
||
name = fn.id
|
||
if name == "rand":
|
||
if len(node.args) != 0:
|
||
raise ExecutionError("rand() takes no arguments")
|
||
return random.random()
|
||
if name == "randint":
|
||
if len(node.args) != 2:
|
||
raise ExecutionError("randint(a,b) requires two arguments")
|
||
a = eval_node(node.args[0])
|
||
b = eval_node(node.args[1])
|
||
try:
|
||
return random.randint(int(a), int(b))
|
||
except Exception as exc: # noqa: BLE001
|
||
raise ExecutionError(f"randint invalid arguments: {exc}")
|
||
if name == "choice":
|
||
if len(node.args) != 1:
|
||
raise ExecutionError("choice(list) requires one argument")
|
||
seq = eval_node(node.args[0])
|
||
if not isinstance(seq, (list, tuple)):
|
||
raise ExecutionError("choice() expects list or tuple")
|
||
if not seq:
|
||
raise ExecutionError("choice() on empty sequence")
|
||
return random.choice(seq)
|
||
raise ExecutionError(f"Function {name} is not allowed")
|
||
# Запрещаем всё остальное (Name/Attribute/Subscript/IfExp/Comprehensions и пр.)
|
||
raise ExecutionError("Expression not allowed")
|
||
|
||
try:
|
||
tree = ast.parse(str(expr), mode="eval")
|
||
except Exception as exc:
|
||
raise ExecutionError(f"SetVars expr parse error: {exc}") from exc
|
||
return eval_node(tree)
|
||
|
||
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
|
||
out_map = context.get("OUT") or {}
|
||
result: Dict[str, Any] = {}
|
||
import re as _re
|
||
for v in sorted(self._normalize(), key=lambda x: x.get("order", 0)):
|
||
name = v.get("name") or ""
|
||
if not _re.fullmatch(r"[A-Za-z_][A-Za-z0-9_]*", name or ""):
|
||
raise ExecutionError(f"SetVars invalid variable name: {name!r}")
|
||
mode = v.get("mode", "string")
|
||
raw_val = v.get("value", "")
|
||
if mode == "expr":
|
||
resolved = self._safe_eval_expr(str(raw_val))
|
||
else:
|
||
resolved = render_template_simple(str(raw_val or ""), context, out_map)
|
||
result[name] = resolved
|
||
return {"vars": result}
|
||
|
||
class ProviderCallNode(Node):
|
||
type_name = "ProviderCall"
|
||
|
||
# --- Prompt Manager helpers -------------------------------------------------
|
||
def _get_blocks(self) -> List[Dict[str, Any]]:
|
||
"""Return normalized list of prompt blocks from config."""
|
||
raw = self.config.get("blocks") or self.config.get("prompt_blocks") or []
|
||
if not isinstance(raw, list):
|
||
return []
|
||
norm: List[Dict[str, Any]] = []
|
||
for i, b in enumerate(raw):
|
||
if not isinstance(b, dict):
|
||
continue
|
||
role = str(b.get("role", "user")).lower().strip()
|
||
if role not in {"system", "user", "assistant", "tool"}:
|
||
role = "user"
|
||
# order fallback: keep original index if not provided/correct
|
||
try:
|
||
order = int(b.get("order")) if b.get("order") is not None else i
|
||
except Exception: # noqa: BLE001
|
||
order = i
|
||
norm.append(
|
||
{
|
||
"id": b.get("id") or f"b{i}",
|
||
"name": b.get("name") or f"Block {i+1}",
|
||
"role": role,
|
||
"prompt": b.get("prompt") or "",
|
||
"enabled": bool(b.get("enabled", True)),
|
||
"order": order,
|
||
}
|
||
)
|
||
return norm
|
||
|
||
def _render_blocks_to_unified(self, context: Dict[str, Any]) -> List[Dict[str, Any]]:
|
||
"""
|
||
Filter+sort+render blocks to unified messages:
|
||
[{role, content, name?}]
|
||
"""
|
||
out_map = context.get("OUT") or {}
|
||
blocks = [b for b in self._get_blocks() if b.get("enabled", True)]
|
||
blocks.sort(key=lambda x: x.get("order", 0))
|
||
messages: List[Dict[str, Any]] = []
|
||
for b in blocks:
|
||
content = render_template_simple(str(b.get("prompt") or ""), context, out_map)
|
||
# name поля блоков не передаются в провайдерские payload'ы
|
||
msg = {"role": b["role"], "content": content}
|
||
messages.append(msg)
|
||
return messages
|
||
|
||
def _messages_to_payload(self, provider: str, messages: List[Dict[str, Any]], context: Dict[str, Any]) -> Dict[str, Any]:
|
||
"""Convert unified messages to provider-specific request payload."""
|
||
params = context.get("params") or {}
|
||
model = context.get("model") or ""
|
||
if provider == "openai":
|
||
payload: Dict[str, Any] = {
|
||
"model": model,
|
||
"messages": [
|
||
{"role": m["role"], "content": m["content"]}
|
||
for m in messages
|
||
],
|
||
"temperature": params.get("temperature", 0.7),
|
||
}
|
||
if params.get("max_tokens") is not None:
|
||
payload["max_tokens"] = params.get("max_tokens")
|
||
if params.get("top_p") is not None:
|
||
payload["top_p"] = params.get("top_p")
|
||
if params.get("stop") is not None:
|
||
payload["stop"] = params.get("stop")
|
||
return payload
|
||
|
||
if provider == "gemini":
|
||
sys_text = "\n\n".join([m["content"] for m in messages if m["role"] == "system"]).strip()
|
||
contents = []
|
||
for m in messages:
|
||
if m["role"] == "system":
|
||
continue
|
||
role = "model" if m["role"] == "assistant" else "user"
|
||
contents.append({"role": role, "parts": [{"text": m["content"]}]})
|
||
gen_cfg: Dict[str, Any] = {}
|
||
if params.get("temperature") is not None:
|
||
gen_cfg["temperature"] = params.get("temperature")
|
||
if params.get("max_tokens") is not None:
|
||
gen_cfg["maxOutputTokens"] = params.get("max_tokens")
|
||
if params.get("top_p") is not None:
|
||
gen_cfg["topP"] = params.get("top_p")
|
||
if params.get("stop") is not None:
|
||
gen_cfg["stopSequences"] = params.get("stop")
|
||
payload = {"model": model, "contents": contents}
|
||
if sys_text:
|
||
payload["systemInstruction"] = {"parts": [{"text": sys_text}]}
|
||
if gen_cfg:
|
||
payload["generationConfig"] = gen_cfg
|
||
return payload
|
||
|
||
if provider == "claude":
|
||
sys_text = "\n\n".join([m["content"] for m in messages if m["role"] == "system"]).strip()
|
||
msgs = []
|
||
for m in messages:
|
||
if m["role"] == "system":
|
||
continue
|
||
role = m["role"] if m["role"] in {"user", "assistant"} else "user"
|
||
msgs.append({"role": role, "content": [{"type": "text", "text": m["content"]}]})
|
||
payload: Dict[str, Any] = {
|
||
"model": model,
|
||
"messages": msgs,
|
||
}
|
||
if sys_text:
|
||
payload["system"] = sys_text
|
||
if params.get("temperature") is not None:
|
||
payload["temperature"] = params.get("temperature")
|
||
if params.get("max_tokens") is not None:
|
||
payload["max_tokens"] = params.get("max_tokens")
|
||
if params.get("top_p") is not None:
|
||
payload["top_p"] = params.get("top_p")
|
||
if params.get("stop") is not None:
|
||
payload["stop"] = params.get("stop")
|
||
return payload
|
||
|
||
return {}
|
||
|
||
def _blocks_struct_for_template(self, provider: str, messages: List[Dict[str, Any]], context: Dict[str, Any]) -> Dict[str, Any]:
|
||
"""
|
||
Сформировать структуру для вставки в шаблон (template) из Prompt Blocks.
|
||
Возвращает provider-специфичные ключи, которые можно вставлять в JSON:
|
||
- openai: { "messages": [...] , "system_text": "..." }
|
||
- gemini: { "contents": [...], "systemInstruction": {...}, "system_text": "..." }
|
||
- claude: { "system_text": "...", "system": "...", "messages": [...] }
|
||
"""
|
||
provider = (provider or "openai").lower()
|
||
# Гарантируем список
|
||
msgs = messages or []
|
||
|
||
if provider == "openai":
|
||
# Уже в формате {"role","content"}
|
||
sys_text = "\n\n".join([m["content"] for m in msgs if m.get("role") == "system"]).strip()
|
||
# Вставляем как есть (editor будет встраивать JSON массива без кавычек)
|
||
return {
|
||
"messages": [
|
||
{"role": m["role"], "content": m.get("content")}
|
||
for m in msgs
|
||
],
|
||
"system_text": sys_text,
|
||
}
|
||
|
||
if provider == "gemini":
|
||
sys_text = "\n\n".join([m["content"] for m in msgs if m.get("role") == "system"]).strip()
|
||
contents = []
|
||
for m in msgs:
|
||
if m.get("role") == "system":
|
||
continue
|
||
role = "model" if m.get("role") == "assistant" else "user"
|
||
contents.append({"role": role, "parts": [{"text": str(m.get("content") or "")}]})
|
||
d: Dict[str, Any] = {
|
||
"contents": contents,
|
||
"system_text": sys_text,
|
||
}
|
||
if sys_text:
|
||
d["systemInstruction"] = {"parts": [{"text": sys_text}]}
|
||
return d
|
||
|
||
if provider == "claude":
|
||
sys_text = "\n\n".join([m["content"] for m in msgs if m.get("role") == "system"]).strip()
|
||
out_msgs = []
|
||
for m in msgs:
|
||
if m.get("role") == "system":
|
||
continue
|
||
role = m.get("role")
|
||
role = role if role in {"user", "assistant"} else "user"
|
||
out_msgs.append({"role": role, "content": [{"type": "text", "text": str(m.get("content") or "")}]})
|
||
return {
|
||
"system_text": sys_text,
|
||
"system": sys_text, # удобно для шаблона: "system": "{{ pm.system_text }}"
|
||
"messages": out_msgs,
|
||
}
|
||
|
||
# По умолчанию ничего, но это валидный JSON
|
||
return {"messages": []}
|
||
|
||
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
|
||
provider = (self.config.get("provider") or "openai").lower()
|
||
|
||
# Support provider-specific configs stored in UI as provider_configs.{provider}
|
||
prov_cfg: Dict[str, Any] = {}
|
||
try:
|
||
cfgs = self.config.get("provider_configs") or {}
|
||
if isinstance(cfgs, dict):
|
||
prov_cfg = cfgs.get(provider) or {}
|
||
except Exception: # noqa: BLE001
|
||
prov_cfg = {}
|
||
|
||
base_url = prov_cfg.get("base_url") or self.config.get("base_url")
|
||
if not base_url:
|
||
raise ExecutionError(f"Node {self.node_id} ({self.type_name}) requires 'base_url' in config")
|
||
if not str(base_url).startswith(("http://", "https://")):
|
||
base_url = "http://" + str(base_url)
|
||
|
||
endpoint_tmpl: str = prov_cfg.get("endpoint") or self.config.get("endpoint") or ""
|
||
template: str = prov_cfg.get("template") or self.config.get("template") or "{}"
|
||
headers_json: str = prov_cfg.get("headers") or self.config.get("headers") or "{}"
|
||
|
||
# Default endpoints if not set
|
||
if not endpoint_tmpl:
|
||
if provider == "openai":
|
||
endpoint_tmpl = "/v1/chat/completions"
|
||
elif provider == "gemini":
|
||
endpoint_tmpl = "/v1beta/models/{{ model }}:generateContent"
|
||
elif provider == "claude":
|
||
endpoint_tmpl = "/v1/messages"
|
||
|
||
# Подготовим Prompt Blocks + pm-структуру для шаблона
|
||
unified_msgs = self._render_blocks_to_unified(context)
|
||
pm_struct = self._blocks_struct_for_template(provider, unified_msgs, context)
|
||
|
||
# Расширяем контекст для рендера шаблонов
|
||
render_ctx = dict(context)
|
||
render_ctx["pm"] = pm_struct
|
||
|
||
# Единый JSON-фрагмент PROMPT для шаблонов: [[PROMPT]]
|
||
prompt_fragment = ""
|
||
try:
|
||
if provider == "openai":
|
||
prompt_fragment = '"messages": ' + json.dumps(pm_struct.get("messages", []), ensure_ascii=False)
|
||
elif provider == "gemini":
|
||
parts = []
|
||
contents = pm_struct.get("contents")
|
||
if contents is not None:
|
||
parts.append('"contents": ' + json.dumps(contents, ensure_ascii=False))
|
||
sysi = pm_struct.get("systemInstruction")
|
||
# даже если пустой объект {}, это валидно
|
||
if sysi is not None:
|
||
parts.append('"systemInstruction": ' + json.dumps(sysi, ensure_ascii=False))
|
||
prompt_fragment = ", ".join(parts)
|
||
elif provider == "claude":
|
||
parts = []
|
||
sys_text = pm_struct.get("system_text") or pm_struct.get("system")
|
||
if sys_text is not None:
|
||
parts.append('"system": ' + json.dumps(sys_text, ensure_ascii=False))
|
||
msgs = pm_struct.get("messages")
|
||
if msgs is not None:
|
||
parts.append('"messages": ' + json.dumps(msgs, ensure_ascii=False))
|
||
prompt_fragment = ", ".join(parts)
|
||
except Exception: # noqa: BLE001
|
||
prompt_fragment = ""
|
||
render_ctx["PROMPT"] = prompt_fragment
|
||
|
||
# Render helper с поддержкой [[VAR]], [[OUT]] и {{ ... }}
|
||
def render(s: str) -> str:
|
||
return render_template_simple(s or "", render_ctx, render_ctx.get("OUT") or {})
|
||
|
||
# Рендер endpoint с макросами/шаблонами
|
||
endpoint = render(endpoint_tmpl)
|
||
|
||
# Формируем тело ТОЛЬКО из template/[[PROMPT]] (без сырого payload/входов).
|
||
# Больше НИКАКОГО фоллбэка на unified-построение: если шаблон невалиден — это ошибка ноды.
|
||
try:
|
||
rendered = render(template)
|
||
# DEBUG: печать отрендеренного шаблона с номерами строк для точной диагностики JSONDecodeError
|
||
try:
|
||
_lines = rendered.splitlines()
|
||
_preview = "\n".join(f"{i+1:03d}: {_lines[i]}" for i in range(min(len(_lines), 120)))
|
||
print(f"DEBUG: ProviderCallNode rendered_template node={self.node_id} provider={provider}\\n{_preview}")
|
||
except Exception:
|
||
try:
|
||
print(f"DEBUG: ProviderCallNode rendered_template(node={self.node_id}, provider={provider}) len={len(rendered)}")
|
||
except Exception:
|
||
pass
|
||
payload = json.loads(rendered)
|
||
except Exception as exc: # noqa: BLE001
|
||
raise ExecutionError(f"ProviderCall template invalid JSON: {exc}")
|
||
|
||
# Заголовки — полностью из редактируемого JSON с макросами
|
||
try:
|
||
headers_src = render(headers_json) if headers_json else "{}"
|
||
headers = json.loads(headers_src) if headers_src else {}
|
||
if not isinstance(headers, dict):
|
||
raise ValueError("headers must be a JSON object")
|
||
except Exception as exc: # noqa: BLE001
|
||
raise ExecutionError(f"ProviderCall headers invalid JSON: {exc}")
|
||
|
||
# Итоговый URL
|
||
if not base_url.startswith(("http://", "https://")):
|
||
base_url = "http://" + base_url
|
||
url = endpoint if endpoint.startswith("http") else urljoin(base_url.rstrip('/') + '/', endpoint.lstrip('/'))
|
||
|
||
# Debug logs to validate config selection and payload
|
||
# Brute request/response logging (FULL, no masking)
|
||
try:
|
||
final_headers = {"Content-Type": "application/json", **headers}
|
||
print("===== ProviderCall REQUEST BEGIN =====")
|
||
print(f"node={self.node_id} type={self.type_name} provider={provider}")
|
||
print(f"URL: {url}")
|
||
try:
|
||
print("Headers:")
|
||
print(json.dumps(final_headers, ensure_ascii=False, indent=2))
|
||
except Exception:
|
||
print(f"Headers(raw): {final_headers}")
|
||
try:
|
||
print("Body JSON:")
|
||
print(json.dumps(payload, ensure_ascii=False, indent=2))
|
||
except Exception:
|
||
print(f"Body(raw): {payload}")
|
||
print("===== ProviderCall REQUEST END =====")
|
||
except Exception:
|
||
pass
|
||
|
||
async with build_client() as client:
|
||
resp = await client.post(url, json=payload, headers=final_headers)
|
||
# Do not raise_for_status: keep body/logs on 4xx/5xx
|
||
try:
|
||
print("===== ProviderCall RESPONSE BEGIN =====")
|
||
print(f"node={self.node_id} type={self.type_name} provider={provider}")
|
||
print(f"Status: {resp.status_code}")
|
||
try:
|
||
print("Headers:")
|
||
print(json.dumps(dict(resp.headers), ensure_ascii=False, indent=2))
|
||
except Exception:
|
||
try:
|
||
print(f"Headers(raw): {dict(resp.headers)}")
|
||
except Exception:
|
||
print("Headers(raw): <unavailable>")
|
||
try:
|
||
body_text = resp.text
|
||
except Exception:
|
||
body_text = "<resp.text decode error>"
|
||
print("Body Text:")
|
||
print(body_text)
|
||
print("===== ProviderCall RESPONSE END =====")
|
||
except Exception:
|
||
pass
|
||
|
||
try:
|
||
data = resp.json()
|
||
except Exception:
|
||
data = {"error": "Failed to decode JSON from upstream", "text": resp.text}
|
||
|
||
# Извлекаем текст best-effort
|
||
text = None
|
||
if provider == "openai":
|
||
try:
|
||
text = data.get("choices", [{}])[0].get("message", {}).get("content")
|
||
except Exception: # noqa: BLE001
|
||
text = None
|
||
elif provider == "gemini":
|
||
try:
|
||
text = data.get("candidates", [{}])[0].get("content", {}).get("parts", [{}])[0].get("text")
|
||
except Exception: # noqa: BLE001
|
||
text = None
|
||
elif provider == "claude":
|
||
try:
|
||
blocks = data.get("content") or []
|
||
texts = [b.get("text") for b in blocks if isinstance(b, dict) and b.get("type") == "text"]
|
||
text = "\n".join([t for t in texts if isinstance(t, str)])
|
||
except Exception: # noqa: BLE001
|
||
text = None
|
||
|
||
return {"result": data, "response_text": text or ""}
|
||
|
||
|
||
class RawForwardNode(Node):
|
||
type_name = "RawForward"
|
||
|
||
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
|
||
incoming = context.get("incoming", {})
|
||
raw_payload = incoming.get("json")
|
||
|
||
base_url: Optional[str] = self.config.get("base_url")
|
||
override_path: Optional[str] = self.config.get("override_path")
|
||
|
||
# Разрешаем макросы в конфиге RawForward (base_url, override_path)
|
||
out_map_for_macros = context.get("OUT") or {}
|
||
if base_url:
|
||
base_url = render_template_simple(str(base_url), context, out_map_for_macros)
|
||
if override_path:
|
||
override_path = render_template_simple(str(override_path), context, out_map_for_macros)
|
||
|
||
# Если base_url не указан, включаем автодетекцию
|
||
if not base_url:
|
||
vendor = detect_vendor(raw_payload)
|
||
if vendor == "openai":
|
||
base_url = "https://api.openai.com"
|
||
elif vendor == "claude":
|
||
base_url = "https://api.anthropic.com"
|
||
elif vendor == "gemini":
|
||
base_url = "https://generativelanguage.googleapis.com"
|
||
else:
|
||
raise ExecutionError(f"Node {self.node_id} ({self.type_name}): 'base_url' is not configured and vendor could not be detected.")
|
||
|
||
# Гарантируем наличие схемы у base_url
|
||
if not base_url.startswith(("http://", "https://")):
|
||
base_url = "http://" + base_url
|
||
|
||
path = override_path or incoming.get("path") or "/"
|
||
query = incoming.get("query")
|
||
|
||
if query:
|
||
path_with_qs = f"{path}?{query}"
|
||
else:
|
||
path_with_qs = path
|
||
|
||
url = urljoin(base_url.rstrip("/") + "/", path_with_qs.lstrip("/"))
|
||
|
||
passthrough_headers: bool = bool(self.config.get("passthrough_headers", True))
|
||
extra_headers_json: str = self.config.get("extra_headers") or "{}"
|
||
# Макросы в extra_headers
|
||
try:
|
||
extra_headers_src = render_template_simple(extra_headers_json, context, out_map_for_macros) if extra_headers_json else "{}"
|
||
extra_headers = json.loads(extra_headers_src) if extra_headers_src else {}
|
||
if not isinstance(extra_headers, dict):
|
||
raise ValueError("extra_headers must be an object")
|
||
except Exception as exc: # noqa: BLE001
|
||
raise ExecutionError(f"RawForward extra_headers invalid JSON: {exc}")
|
||
|
||
headers: Dict[str, str] = {}
|
||
if passthrough_headers:
|
||
inc_headers = incoming.get("headers") or {}
|
||
# Копируем все заголовки, кроме Host и Content-Length
|
||
for k, v in inc_headers.items():
|
||
if k.lower() not in ['host', 'content-length']:
|
||
headers[k] = v
|
||
|
||
# Убедимся, что Content-Type на месте, если его не было
|
||
if 'content-type' not in {k.lower() for k in headers}:
|
||
headers['Content-Type'] = 'application/json'
|
||
|
||
headers.update(extra_headers)
|
||
|
||
# Brute request/response logging (FULL, no masking)
|
||
try:
|
||
print("===== RawForward REQUEST BEGIN =====")
|
||
print(f"node={self.node_id} type={self.type_name}")
|
||
print(f"URL: {url}")
|
||
try:
|
||
print("Headers:")
|
||
print(json.dumps(headers, ensure_ascii=False, indent=2))
|
||
except Exception:
|
||
print(f"Headers(raw): {headers}")
|
||
try:
|
||
print("Body JSON:")
|
||
print(json.dumps(raw_payload, ensure_ascii=False, indent=2))
|
||
except Exception:
|
||
print(f"Body(raw): {raw_payload}")
|
||
print("===== RawForward REQUEST END =====")
|
||
except Exception:
|
||
pass
|
||
|
||
async with build_client() as client:
|
||
resp = await client.post(url, json=raw_payload, headers=headers)
|
||
|
||
# Response logging
|
||
try:
|
||
print("===== RawForward RESPONSE BEGIN =====")
|
||
print(f"node={self.node_id} type={self.type_name}")
|
||
print(f"Status: {resp.status_code}")
|
||
try:
|
||
print("Headers:")
|
||
print(json.dumps(dict(resp.headers), ensure_ascii=False, indent=2))
|
||
except Exception:
|
||
try:
|
||
print(f"Headers(raw): {dict(resp.headers)}")
|
||
except Exception:
|
||
print("Headers(raw): <unavailable>")
|
||
try:
|
||
body_text = resp.text
|
||
except Exception:
|
||
body_text = "<resp.text decode error>"
|
||
print("Body Text:")
|
||
print(body_text)
|
||
print("===== RawForward RESPONSE END =====")
|
||
except Exception:
|
||
pass
|
||
|
||
# Decode JSON if possible, otherwise return text
|
||
try:
|
||
data = resp.json()
|
||
except Exception:
|
||
data = {"error": "Failed to decode JSON from upstream", "text": resp.text}
|
||
return {"result": data}
|
||
|
||
|
||
class ReturnNode(Node):
|
||
type_name = "Return"
|
||
|
||
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
|
||
# Определяем целевой формат
|
||
cfg = self.config or {}
|
||
target = str(cfg.get("target_format", "auto")).lower().strip() or "auto"
|
||
if target == "auto":
|
||
target = str(context.get("vendor_format") or "openai").lower().strip() or "openai"
|
||
|
||
# Рендерим текст из шаблона (по умолчанию берём [[OUT1]])
|
||
out_map = context.get("OUT") or {}
|
||
template = cfg.get("text_template")
|
||
if template is None or template == "":
|
||
template = "[[OUT1]]"
|
||
try:
|
||
text = render_template_simple(str(template), context, out_map)
|
||
except Exception:
|
||
text = ""
|
||
|
||
model = str(context.get("model") or "")
|
||
|
||
# Форматтеры под провайдеры (как в execute_pipeline_echo)
|
||
def fmt_openai(t: str) -> Dict[str, Any]:
|
||
return {
|
||
"id": "ret_mock_123",
|
||
"object": "chat.completion",
|
||
"model": model,
|
||
"choices": [
|
||
{
|
||
"index": 0,
|
||
"message": {"role": "assistant", "content": t},
|
||
"finish_reason": "stop",
|
||
}
|
||
],
|
||
"usage": {"prompt_tokens": 0, "completion_tokens": len((t or "").split()), "total_tokens": 0},
|
||
}
|
||
|
||
def fmt_gemini(t: str) -> Dict[str, Any]:
|
||
return {
|
||
"candidates": [
|
||
{
|
||
"content": {
|
||
"role": "model",
|
||
"parts": [{"text": t}],
|
||
},
|
||
"finishReason": "STOP",
|
||
"index": 0,
|
||
}
|
||
],
|
||
"modelVersion": model,
|
||
}
|
||
|
||
def fmt_claude(t: str) -> Dict[str, Any]:
|
||
return {
|
||
"id": "msg_ret_123",
|
||
"type": "message",
|
||
"model": model,
|
||
"role": "assistant",
|
||
"content": [
|
||
{"type": "text", "text": t}
|
||
],
|
||
"stop_reason": "end_turn",
|
||
}
|
||
|
||
if target == "openai":
|
||
result = fmt_openai(text)
|
||
elif target == "gemini":
|
||
result = fmt_gemini(text)
|
||
elif target == "claude":
|
||
result = fmt_claude(text)
|
||
else:
|
||
# неизвестное значение — безопасный дефолт
|
||
result = fmt_openai(text)
|
||
|
||
return {"result": result, "response_text": text}
|
||
|
||
|
||
|
||
NODE_REGISTRY.update({
|
||
SetVarsNode.type_name: SetVarsNode,
|
||
ProviderCallNode.type_name: ProviderCallNode,
|
||
RawForwardNode.type_name: RawForwardNode,
|
||
ReturnNode.type_name: ReturnNode,
|
||
})
|
||
|
||
|