AI Foundations

Overview

Provider-independent AI orchestration in quantsmind.ai: messages, model-backend interface, conversation memory, tools, evaluation metrics, structured-output helpers, and a minimal agent loop. Model weights and provider integrations live outside this package by design.

Purpose

Let users build model-agnostic agents, wire tools, and evaluate outputs without depending on any commercial AI provider.

Concept

ModelBackend.complete() is the single seam to any language model. Everything else — ConversationMemory windowing, ToolRegistry dispatch, Agent plan-act-observe loop, exact_match/token_f1 metrics, JSON output parsing — is deterministic SDK logic.

API

Message (+ role constants), ModelBackend, ModelCapabilities, ConversationMemory, Tool, ToolCall, ToolRegistry, UnknownToolError, exact_match(), token_f1(), parse_json_output(), Agent, AgentResult, TOOL_CALL_KEY.

Input / Processing / Output

Input: task strings and tool calls. Processing: bounded loop with memory recall. Output: AgentResult(answer, steps, tool_calls, finished).

Example

python examples/ai/tool_loop.py answers (2+3)*4 through two dispatched tool calls with a labeled scripted backend.

Limitations

No model weights, no provider integrations, no streaming, no embeddings; the scripted backends in examples/tests are scaffolding, not inference.