Add OLLAMA_NUM_CTX config to reduce VRAM usage

Replaces the hardcoded num_ctx=32768 KV cache allocation with a
configurable env var defaulting to 8192. This significantly reduces
VRAM pressure when multiple services share the GPU.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

src/fabledassistant/config.py

@@ -27,6 +27,9 @@ class Config:
     # Optional dedicated model for intent classification (should be small/fast).
     # Falls back to OLLAMA_MODEL if not set. Can also be overridden per-user via settings.
     OLLAMA_INTENT_MODEL: str = os.environ.get("OLLAMA_INTENT_MODEL", "")
+    # KV cache context window for generation. Lower = less VRAM, less throughput impact.
+    # 8192 is sufficient for most conversations; raise if you paste large documents.
+    OLLAMA_NUM_CTX: int = int(os.environ.get("OLLAMA_NUM_CTX", "8192"))
     SECRET_KEY: str = _read_secret("SECRET_KEY", "SECRET_KEY_FILE", "dev-secret-change-me")
     SECURE_COOKIES: bool = os.environ.get("SECURE_COOKIES", "").lower() in ("1", "true", "yes")
     LOG_LEVEL: str = os.environ.get("LOG_LEVEL", "INFO")

src/fabledassistant/services/llm.py

@@ -80,7 +80,7 @@ async def stream_chat(
     options: dict | None = None,
 ) -> AsyncGenerator[str, None]:
     """Stream chat completion from Ollama, yielding content chunks."""
-    merged_options = {"num_ctx": 32768}
+    merged_options = {"num_ctx": Config.OLLAMA_NUM_CTX}
     if options:
         merged_options.update(options)
     payload: dict = {"model": model, "messages": messages, "stream": True, "options": merged_options}
@@ -126,7 +126,7 @@ async def stream_chat_with_tools(
     Thinking tokens are consumed by Ollama and not forwarded to the caller;
     only the final response content is yielded. Expect higher TTFT when enabled.
     """
-    options: dict = {"num_ctx": 32768}
+    options: dict = {"num_ctx": Config.OLLAMA_NUM_CTX}
     if tools:
         options["num_predict"] = 8192
     payload: dict = {