prompt tweaks

* max-height and overflow

* max-height and overflow

.github/workflows/ci.yml

@@ -19,25 +19,6 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - name: Remove unnecessary files to release disk space
-        run: |
-          sudo rm -rf \
-            "$AGENT_TOOLSDIRECTORY" \
-            /opt/ghc \
-            /opt/google/chrome \
-            /opt/microsoft/msedge \
-            /opt/microsoft/powershell \
-            /opt/pipx \
-            /usr/lib/mono \
-            /usr/local/julia* \
-            /usr/local/lib/android \
-            /usr/local/lib/node_modules \
-            /usr/local/share/chromium \
-            /usr/local/share/powershell \
-            /usr/local/share/powershell \
-            /usr/share/dotnet \
-            /usr/share/swift
-
       - name: Log in to GHCR
         uses: docker/login-action@v3
         with:

.github/workflows/test-container-build.yml

@@ -14,25 +14,6 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - name: Remove unnecessary files to release disk space
-        run: |
-          sudo rm -rf \
-            "$AGENT_TOOLSDIRECTORY" \
-            /opt/ghc \
-            /opt/google/chrome \
-            /opt/microsoft/msedge \
-            /opt/microsoft/powershell \
-            /opt/pipx \
-            /usr/lib/mono \
-            /usr/local/julia* \
-            /usr/local/lib/android \
-            /usr/local/lib/node_modules \
-            /usr/local/share/chromium \
-            /usr/local/share/powershell \
-            /usr/local/share/powershell \
-            /usr/share/dotnet \
-            /usr/share/swift
-
       - name: Log in to GHCR
         uses: docker/login-action@v3
         with:

.gitignore

@@ -31,5 +31,4 @@ scenes/
 !scenes/infinity-quest/infinity-quest.json
 tts_voice_samples/*.wav
 third-party-docs/
-legacy-state-reinforcements.yaml
-CLAUDE.md
+legacy-state-reinforcements.yaml

Dockerfile

@@ -45,9 +45,6 @@ WORKDIR /app
 
 RUN apt-get update && apt-get install -y \
     bash \
-    wget \
-    tar \
-    xz-utils \
     && rm -rf /var/lib/apt/lists/*
 
 # Install uv in the final stage
@@ -56,21 +53,6 @@ RUN pip install uv
 # Copy virtual environment from backend-build stage
 COPY --from=backend-build /app/.venv /app/.venv
 
-# Download and install FFmpeg 8.0 with shared libraries into .venv (matching Windows installer approach)
-# Using BtbN FFmpeg builds which provide shared libraries - verified to work
-# Note: We tried using jrottenberg/ffmpeg:8.0-ubuntu image but copying libraries from it didn't work properly,
-#       so we use the direct download approach which is more reliable and matches the Windows installer
-RUN cd /tmp && \
-    wget -q https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-master-latest-linux64-gpl-shared.tar.xz -O ffmpeg.tar.xz && \
-    tar -xf ffmpeg.tar.xz && \
-    cp -a ffmpeg-master-latest-linux64-gpl-shared/bin/* /app/.venv/bin/ && \
-    cp -a ffmpeg-master-latest-linux64-gpl-shared/lib/* /app/.venv/lib/ && \
-    rm -rf ffmpeg-master-latest-linux64-gpl-shared ffmpeg.tar.xz && \
-    LD_LIBRARY_PATH=/app/.venv/lib /app/.venv/bin/ffmpeg -version | head -n 1
-
-# Set LD_LIBRARY_PATH so torchcodec can find ffmpeg libraries at runtime
-ENV LD_LIBRARY_PATH=/app/.venv/lib:${LD_LIBRARY_PATH}
-
 # Copy Python source code
 COPY --from=backend-build /app/src /app/src
 

docs/dev/client/example/runpod_vllm/__init__.py

@@ -0,0 +1,134 @@
+"""
+An attempt to write a client against the runpod serverless vllm worker.
+
+This is close to functional, but since runpod serverless gpu availability is currently terrible, i have
+been unable to properly test it.
+
+Putting it here for now since i think it makes a decent example of how to write a client against a new service.
+"""
+
+import pydantic
+import structlog
+import runpod
+import asyncio
+import aiohttp
+from talemate.client.base import ClientBase, ExtraField
+from talemate.client.registry import register
+from talemate.emit import emit
+from talemate.config import Client as BaseClientConfig
+
+log = structlog.get_logger("talemate.client.runpod_vllm")
+
+
+class Defaults(pydantic.BaseModel):
+    max_token_length: int = 4096
+    model: str = ""
+    runpod_id: str = ""
+
+
+class ClientConfig(BaseClientConfig):
+    runpod_id: str = ""
+
+
+@register()
+class RunPodVLLMClient(ClientBase):
+    client_type = "runpod_vllm"
+    conversation_retries = 5
+    config_cls = ClientConfig
+
+    class Meta(ClientBase.Meta):
+        title: str = "Runpod VLLM"
+        name_prefix: str = "Runpod VLLM"
+        enable_api_auth: bool = True
+        manual_model: bool = True
+        defaults: Defaults = Defaults()
+        extra_fields: dict[str, ExtraField] = {
+            "runpod_id": ExtraField(
+                name="runpod_id",
+                type="text",
+                label="Runpod ID",
+                required=True,
+                description="The Runpod ID to connect to.",
+            )
+        }
+
+    def __init__(self, model=None, runpod_id=None, **kwargs):
+        self.model_name = model
+        self.runpod_id = runpod_id
+        super().__init__(**kwargs)
+
+    @property
+    def experimental(self):
+        return False
+
+    def set_client(self, **kwargs):
+        log.debug("set_client", kwargs=kwargs, runpod_id=self.runpod_id)
+        self.runpod_id = kwargs.get("runpod_id", self.runpod_id)
+
+    def tune_prompt_parameters(self, parameters: dict, kind: str):
+        super().tune_prompt_parameters(parameters, kind)
+
+        keys = list(parameters.keys())
+
+        valid_keys = ["temperature", "top_p", "max_tokens"]
+
+        for key in keys:
+            if key not in valid_keys:
+                del parameters[key]
+
+    async def get_model_name(self):
+        return self.model_name
+
+    async def generate(self, prompt: str, parameters: dict, kind: str):
+        """
+        Generates text from the given prompt and parameters.
+        """
+        prompt = prompt.strip()
+
+        self.log.debug("generate", prompt=prompt[:128] + " ...", parameters=parameters)
+
+        try:
+            async with aiohttp.ClientSession() as session:
+                endpoint = runpod.AsyncioEndpoint(self.runpod_id, session)
+
+                run_request = await endpoint.run(
+                    {
+                        "input": {
+                            "prompt": prompt,
+                        }
+                        # "parameters": parameters
+                    }
+                )
+
+                while (await run_request.status()) not in [
+                    "COMPLETED",
+                    "FAILED",
+                    "CANCELLED",
+                ]:
+                    status = await run_request.status()
+                    log.debug("generate", status=status)
+                    await asyncio.sleep(0.1)
+
+                status = await run_request.status()
+
+                log.debug("generate", status=status)
+
+                response = await run_request.output()
+
+                log.debug("generate", response=response)
+
+                return response["choices"][0]["tokens"][0]
+
+        except Exception as e:
+            self.log.error("generate error", e=e)
+            emit(
+                "status", message="Error during generation (check logs)", status="error"
+            )
+            return ""
+
+    def reconfigure(self, **kwargs):
+        if kwargs.get("model"):
+            self.model_name = kwargs["model"]
+        if "runpod_id" in kwargs:
+            self.api_auth = kwargs["runpod_id"]
+        self.set_client(**kwargs)

docs/getting-started/installation/windows.md

@@ -12,6 +12,14 @@
 !!! note "First start can take a while"
     The initial download and dependency installation may take several minutes, especially on slow internet connections. The console will keep you updated – just wait until the Talemate logo shows up.
 
+### Optional: CUDA support
+
+If you have an NVIDIA GPU and want CUDA acceleration for larger embedding models:
+
+1. Close Talemate (if it is running).
+2. Double-click **`install-cuda.bat`**. This script swaps the CPU-only Torch build for the CUDA 12.8 build.
+3. Start Talemate again via **`start.bat`**.
+
 ## Maintenance & advanced usage
 
 | Script | Purpose |