Add dynamic Synthesizer LLM node

src/nodetool/nodes/llms/__init__.py

@@ -0,0 +1,3 @@
+from .synthesizer import Synthesizer
+
+__all__ = ["Synthesizer"]

src/nodetool/nodes/llms/synthesizer.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+
+import re
+from jinja2 import Environment, BaseLoader
+from pydantic import Field
+
+from nodetool.metadata.types import (
+    Message,
+    MessageTextContent,
+    LanguageModel,
+    Provider,
+)
+from nodetool.chat.providers import Chunk
+from nodetool.workflows.base_node import BaseNode
+from nodetool.workflows.processing_context import ProcessingContext
+
+
+class Synthesizer(BaseNode):
+    """Generate text from a Jinja2 prompt using dynamic properties."""
+
+    _is_dynamic = True
+
+    model: LanguageModel = Field(
+        default=LanguageModel(),
+        description="Model to use for generation",
+    )
+    system: str = Field(
+        default="You are a helpful assistant.",
+        description="System prompt for the LLM",
+    )
+    prompt: str = Field(
+        default="",
+        description="Prompt template rendered with dynamic properties",
+    )
+    max_tokens: int = Field(default=4096, ge=1, le=100000)
+
+    @classmethod
+    def get_title(cls) -> str:
+        return "Synthesizer"
+
+    @classmethod
+    def get_basic_fields(cls) -> list[str]:
+        return ["prompt", "model"]
+
+    async def process(self, context: ProcessingContext) -> str:
+        if self.model.provider == Provider.Empty:
+            raise ValueError("Select a model")
+
+        env = Environment(loader=BaseLoader())
+
+        template_str = self.prompt
+        for var in re.findall(r"{{\s*([^|}]+)", template_str):
+            template_str = template_str.replace(var, var.lower())
+
+        template = env.from_string(template_str)
+        properties = {k.lower(): v for k, v in self._dynamic_properties.items()}
+        user_prompt = template.render(**properties)
+
+        messages = [
+            Message(role="system", content=self.system),
+            Message(role="user", content=[MessageTextContent(text=user_prompt)]),
+        ]
+
+        result = ""
+        async for chunk in context.generate_messages(
+            messages=messages,
+            provider=self.model.provider,
+            model=self.model.id,
+            node_id=self.id,
+            max_tokens=self.max_tokens,
+        ):
+            if isinstance(chunk, Chunk):
+                context.post_message(
+                    Chunk(
+                        node_id=self.id,
+                        content=chunk.content,
+                        content_type=chunk.content_type,
+                    )
+                )
+                result += chunk.content
+        return result

tests/nodetool/test_llms.py

@@ -0,0 +1,32 @@
+import pytest
+
+from nodetool.workflows.processing_context import ProcessingContext
+from nodetool.nodes.llms.synthesizer import Synthesizer
+from nodetool.metadata.types import LanguageModel, Provider
+from nodetool.chat.providers import Chunk
+
+
+@pytest.fixture
+def context():
+    return ProcessingContext(user_id="test", auth_token="test")
+
+
+@pytest.mark.asyncio
+async def test_synthesizer_process(context, monkeypatch):
+    node = Synthesizer(
+        prompt="Hello {{ name }}!",
+        model=LanguageModel(provider=Provider.OpenAI, id="gpt"),
+    )
+    node._dynamic_properties = {"name": "Alice"}
+
+    async def fake_generate_messages(**kwargs):
+        messages = kwargs.get("messages")
+        assert messages[1].content[0].text == "Hello Alice!"
+        yield Chunk(content="Hi Alice", content_type="text")
+
+    monkeypatch.setattr(context, "generate_messages", fake_generate_messages)
+
+    result = await node.process(context)
+    assert result == "Hi Alice"
+
+