Add Friday Core LLM scaffold

Author: elfalehed

modules/friday-core/README.md

@@ -0,0 +1,31 @@
+# Friday Core
+
+This module is the starting point for building a Friday-style assistant stack for The Stark Project.
+
+## LLM Direction
+
+The LLM work lives under [src/llm](src/llm). The initial design focuses on a modular assistant architecture with:
+
+- a model interface layer
+- a chat inference service
+- a tool-routing layer for plugins and actions
+- a future training pipeline for fine-tuning and adaptation
+
+## Current Structure
+
+- [src/llm/README.md](src/llm/README.md) — high-level architecture, roadmap, and technical decisions
+- [src/llm/core/model_interface.py](src/llm/core/model_interface.py) — base model and prompt wrapper abstractions
+- [src/llm/inference/chat_service.py](src/llm/inference/chat_service.py) — minimal chat orchestration layer
+- [src/llm/agents/tool_router.py](src/llm/agents/tool_router.py) — tool registration and dispatch
+
+## Recommended Next Steps
+
+1. Add a concrete backend implementation such as a Hugging Face model wrapper.
+2. Connect the assistant to the existing memory subsystem.
+3. Add a small plugin for a useful action like web lookup or command execution.
+4. Introduce a retrieval layer for long-term context.
+5. Build evaluation and safety checks for assistant behavior.
+
+## Vision
+
+The goal is to evolve this module into an assistant that feels like a Tony Stark-style companion: proactive, context-aware, connected to tools, and capable of acting across the Stark ecosystem.

modules/friday-core/src/llm/README.md

@@ -0,0 +1,94 @@
+# Friday Core LLM Workspace
+
+This folder defines the initial architecture for building a custom assistant stack inspired by the Tony Stark / Friday concept: a fast, multimodal, context-aware assistant with tool use, memory, and safety boundaries.
+
+## Goals
+
+- Build a compact, trainable language model foundation for conversational assistance.
+- Add orchestration layers for memory, tools, and action execution.
+- Keep the design modular so it can evolve from local experiments to a production-grade assistant.
+
+## Proposed Structure
+
+- core/: shared tokenizer, config, model interfaces, and runtime utilities.
+- models/: model definitions, checkpoints, and architecture variants.
+- training/: data pipelines, tokenizer training, pretraining and fine-tuning scripts.
+- inference/: serving, batching, streaming, and prompt execution logic.
+- agents/: planner/executor patterns for tool-calling and multi-step reasoning.
+
+## Recommended Infrastructure
+
+### 1. Runtime
+- Python 3.11+
+- PyTorch or JAX for training and inference
+- Hugging Face Transformers for rapid prototyping
+- vLLM or TensorRT-LLM for optimized serving later
+
+### 2. Data and Memory
+- Structured memory store for long-term facts
+- Episodic memory for recent conversations
+- Vector database for semantic retrieval
+- Event bus integration for tool and sensor subscriptions
+
+### 3. Tooling and Services
+- Plugin interface for commands, APIs, and device control
+- Safety policy layer before action execution
+- Logging and observability for prompts, tool calls, and errors
+
+### 4. Deployment
+- Local development first
+- Containerized inference service
+- Optional GPU-backed training environment
+- Edge deployment path for low-latency assistant use
+
+## Phased Roadmap
+
+### Phase 1: Foundations
+- Define the model interface and configuration schema
+- Build tokenizer and prompt templates
+- Create a minimal inference loop
+- Wire the assistant to the existing memory and plugin layers
+
+### Phase 2: Capability Expansion
+- Add retrieval-augmented generation
+- Introduce tool calling and function routing
+- Support multimodal inputs such as voice and visual context
+- Add conversation state management
+
+### Phase 3: Personality and Alignment
+- Fine-tune on domain-specific assistant behavior
+- Add safety policies and refusal handling
+- Improve memory selection and personalization
+- Optimize latency and response quality
+
+### Phase 4: Stark-like Assistant Experience
+- High-speed voice interaction
+- Context-aware proactive suggestions
+- Multi-agent collaboration for planning and execution
+- Deep integration with robotics, dashboards, and hardware tools
+
+## Technical Decisions
+
+### Why a modular architecture?
+A modular design allows you to experiment with model variants without rewriting the assistant runtime.
+
+### Why start with a small foundation model?
+A smaller model is easier to iterate on and is ideal for local development before scaling to larger architectures.
+
+### Why separate training and inference?
+Training and inference have different dependencies, performance characteristics, and deployment constraints.
+
+### Why integrate memory and tools early?
+An assistant feels intelligent when it can recall context and perform actions, not just generate text.
+
+## Suggested First Implementation
+
+1. Create a minimal model wrapper class.
+2. Add a prompt builder for system, user, and tool context.
+3. Connect the LLM to a simple in-memory conversation store.
+4. Add one tool plugin such as a weather lookup or command runner.
+5. Expose a basic chat endpoint.
+
+## Notes
+
+This is an initial blueprint. The long-term ambition is a Friday-like assistant that can reason, remember, act, and coordinate across the Stark ecosystem.

modules/friday-core/src/llm/agents/tool_router.py

@@ -0,0 +1,19 @@
+from typing import Callable, Dict, List
+
+
+class ToolRouter:
+    """Routes tool calls from LLM outputs to plugin handlers."""
+
+    def __init__(self) -> None:
+        self.tools: Dict[str, Callable[..., str]] = {}
+
+    def register(self, name: str, handler: Callable[..., str]) -> None:
+        self.tools[name] = handler
+
+    def route(self, tool_name: str, *args, **kwargs) -> str:
+        if tool_name not in self.tools:
+            raise KeyError(f"Tool '{tool_name}' is not registered")
+        return self.tools[tool_name](*args, **kwargs)
+
+    def list_tools(self) -> List[str]:
+        return sorted(self.tools.keys())

modules/friday-core/src/llm/core/model_interface.py

@@ -0,0 +1,41 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+
+@dataclass
+class ModelConfig:
+    name: str = "friday-mini"
+    max_context_length: int = 4096
+    temperature: float = 0.7
+    top_p: float = 0.95
+    max_new_tokens: int = 512
+
+
+class LLMBackend:
+    """Abstract interface for any LLM backend used by Friday Core."""
+
+    def __init__(self, config: Optional[ModelConfig] = None) -> None:
+        self.config = config or ModelConfig()
+
+    def generate(self, prompt: str, **kwargs: Any) -> str:
+        raise NotImplementedError
+
+    def stream_generate(self, prompt: str, **kwargs: Any):
+        raise NotImplementedError
+
+
+class FridayLLM:
+    """High-level wrapper that will connect the runtime to models, memory, and tools."""
+
+    def __init__(self, backend: LLMBackend) -> None:
+        self.backend = backend
+
+    def chat(self, message: str, history: Optional[List[Dict[str, str]]] = None) -> str:
+        prompt = self._build_prompt(message, history or [])
+        return self.backend.generate(prompt)
+
+    def _build_prompt(self, message: str, history: List[Dict[str, str]]) -> str:
+        conversation = "\n".join(
+            f"{entry['role']}: {entry['content']}" for entry in history
+        )
+        return f"system: You are Friday, an assistant for The Stark Project.\n{conversation}\nuser: {message}\nassistant:"

modules/friday-core/src/llm/inference/chat_service.py

@@ -0,0 +1,17 @@
+from typing import List, Dict
+
+from ..core.model_interface import FridayLLM, LLMBackend
+
+
+class ChatService:
+    """Minimal service wrapper for a Friday-style chat interface."""
+
+    def __init__(self, backend: LLMBackend) -> None:
+        self.llm = FridayLLM(backend)
+        self.history: List[Dict[str, str]] = []
+
+    def respond(self, message: str) -> str:
+        response = self.llm.chat(message, self.history)
+        self.history.append({"role": "user", "content": message})
+        self.history.append({"role": "assistant", "content": response})
+        return response