src/main.rs

//! Selential Core — MoLoRA Inference Engine
//!
//! A Candle-based MoE-like architecture with a quantized Qwen3 base model,
//! hot-swappable LoRA adapters, and a SmolLM2 query router.
//!
//! Usage:
//!   cargo run --release -- interactive
//!   cargo run --release -- prompt "Write a Rust function"

mod config;
mod engine;
mod hashtags;
mod inference;
mod kb;
mod pipeline;
mod router;
mod translator;



#[allow(unused)]
use clap::{Parser, Subcommand};
use std::path::PathBuf;

#[derive(Parser)]
#[command(name = "selential", version, about = "MoLoRA Inference Engine")]
struct Cli {
    #[command(subcommand)]
    command: Commands,

    /// Path to configuration file
    #[arg(short, long, global = true)]
    config: Option<String>,

    /// Path to base Qwen3 model GGUF file
    #[arg(short = 'm', long, global = true)]
    model: Option<PathBuf>,

    /// GPU device ID (-1 for CPU)
    #[arg(short = 'd', long, default_value = "0", global = true)]
    device: i32,
}

#[derive(Subcommand)]
enum Commands {
    /// Interactive chat mode
    Interactive,
    /// Single prompt inference
    Prompt {
        /// User query
        #[arg(required = true)]
        prompt: Vec<String>,
        /// Expert domain
        #[arg(short, long, default_value = "general")]
        expert: String,
    },
    /// List available experts and routing info
    Info,
    /// Reset conversation
    Reset,
}

fn main() -> anyhow::Result<()> {
    // Initialize tracing with env-filter
    tracing_subscriber::fmt()
        .with_env_filter(
            tracing_subscriber::EnvFilter::try_from_default_env().unwrap_or_else(|_| {
                tracing_subscriber::EnvFilter::new("selential=info,candle=warn")
            }),
        )
        .with_target(false)
        .init();

    let cli = Cli::parse();

    // Load configuration
    let mut config = config::Config::load(cli.config.as_deref())?;

    // Override with CLI arguments
    if let Some(model_path) = cli.model {
        config.base_model_path = model_path;
    }
    if cli.device >= 0 {
        config.gpu_device = cli.device;
    }

    match cli.command {
        Commands::Interactive => run_interactive(config)?,
        Commands::Prompt { prompt, expert } => run_prompt(config, &prompt.join(" "), &expert)?,
        Commands::Info => {
            let expert_names: Vec<String> = config.experts.iter().map(|e| e.name.clone()).collect();
            let router = router::Router::new(&expert_names);
            println!("{}", router.routing_info());
            println!("\nConfiguration:");
            println!("  Base model: {:?}", config.base_model_path);
            println!("  Router model: {:?}", config.router_model_path);
            println!("  Adapters dir: {:?}", config.adapters_dir);
            println!("  Knowledge base: {:?}", config.kb_path);
            println!("  GPU device: {}", config.gpu_device);
            println!("  Max gen tokens: {}", config.max_gen_tokens);
            println!("  Max seq len: {}", config.max_seq_len);
        }
        Commands::Reset => {
            println!("To reset, restart selential or use /reset in interactive mode.");
        }
    }

    Ok(())
}

fn run_prompt(config: config::Config, prompt: &str, expert: &str) -> anyhow::Result<()> {
    let mut engine = inference::InferenceEngine::new(config)?;
    let response = if expert != "general" && !expert.is_empty() {
        engine.process_query_with_expert(prompt, Some(expert))
    } else {
        engine.process_query(prompt)
    }?;
    println!("{}", response);
    Ok(())
}

fn run_interactive(config: config::Config) -> anyhow::Result<()> {
    let mut engine = inference::InferenceEngine::new(config)?;

    println!("\n╔══════════════════════════════════════════════════╗");
    println!("║     Selential Core — MoLoRA Engine v2.0         ║");
    println!("╠══════════════════════════════════════════════════╣");
    println!("║  Orchestra routing: hashtags → expert layers     ║");
    println!("║  🏗️  structural  — struct, impl, trait, enum     ║");
    println!("║  🔀 flow_error  — match, result, concurrency    ║");
    println!("║  📁 system_io   — file I/O, collections         ║");
    println!("╠══════════════════════════════════════════════════╣");
    println!("║  /help  /reset  /stats  /orchestra  /exit       ║");
    println!("║  /hashtags <query>  /tags                       ║");
    println!("╚══════════════════════════════════════════════════╝\n");

    loop {
        let mut input = String::new();
        print!("> ");
        use std::io::Write;
        std::io::stdout().flush()?;
        std::io::stdin().read_line(&mut input)?;
        let input = input.trim();

        if input.is_empty() {
            continue;
        }

        match input {
            "/exit" | "/quit" => {
                println!("Goodbye!");
                break;
            }
            "/reset" => {
                engine.reset();
                println!("Conversation reset.");
                continue;
            }
            "/stats" => {
                println!("{}", serde_json::to_string_pretty(&engine.stats())?);
                continue;
            }
            "/orchestra" => {
                println!("{}", "═".repeat(50));
                println!("🎵 Selential 2.0 — Оркестр экспертов");
                println!("{}", "═".repeat(50));
                println!("");
                println!("  🌐 Layer 1: Generalist Core (#70)");
                println!("     Всегда активен — связность, логика, синтаксис");
                println!("     GGUF: generalist_core.gguf (~24 MB)");
                println!("");
                println!("  🎯 Layer 2: Coding Specialists (по хештегам):");
                println!("");
                println!("  🏗️  structural  — #struct #impl #trait #enum");
                println!("      + #164 (architect) + #92 (impl)");
                println!("      GGUF: structural.gguf (~24 MB)");
                println!("");
                println!("  🔀 flow_error  — #match #result #option #error #concurrency");
                println!("      + #116 (match) + #115 (result)");
                println!("      GGUF: flow_error.gguf (~24 MB)");
                println!("");
                println!("  📁 system_io   — #io #file #collections");
                println!("      + #172 (file I/O) + #116 (match/IO)");
                println!("      GGUF: system_io.gguf (~24 MB)");
                println!("");
                println!("  🦀 rust_coding — legacy (backward compat)");
                println!("      GGUF: rust_coding.gguf (~4 MB)");
                println!("");
                println!("  📊 VRAM: ~47 MB на эксперта | Всего: ~24 MB на оркестр");
                println!("  ⚡ t/s: ~8.7 (с LoRA) vs ~9.7 (baseline) — лишь -10%");
                println!("");
                println!("  Active: {}", engine.active_expert());
                continue;
            }
            "/tags" => {
                println!("Available hashtags for routing:");
                println!("  Code: #struct #impl #trait #enum #match #result #option #error");
                println!("  IO:   #io #file #collections #regex");
                println!("  Async:#concurrency #async #thread");
                println!("  Lang: #rust #python #javascript #zig #cpp #golang");
                println!("  Tone: #casual #teaching #formal");
                continue;
            }
            "/help" => {
                println!("Commands:");
                println!("  /help       - Show this help");
                println!("  /reset      - Reset conversation");
                println!("  /stats      - Show session statistics");
                println!("  /orchestra  - Show orchestra routing info");
                println!("  /tags       - List available hashtags");
                println!("  /hashtags   - Extract hashtags from a query");
                println!("  /pipeline   - Show pipeline info (KB, cache)");
                println!("  /exit       - Exit the program");
                println!(
                    "\nActive: {} | Any other input = query",
                    engine.active_expert()
                );
                continue;
            }
            "/pipeline" => {
                println!("{}", engine.pipeline_info());
                continue;
            }
            s if s.starts_with("/hashtags ") => {
                let query = &s["/hashtags ".len()..];
                let tags = hashtags::extract_hashtags(query);
                println!("Query: {}", query);
                println!("Hashtags: {}", tags.join(" "));
                let is_ru = hashtags::is_russian(query);
                println!(
                    "Language: {}",
                    if is_ru { "Russian" } else { "English/mixed" }
                );
                // Show which orchestra would be selected
                let _tag_names: Vec<String> = tags
                    .iter()
                    .map(|t| t.trim_start_matches('#').to_string())
                    .collect();
                println!("Would route to: (tag-based orchestra detection)");
                continue;
            }
            _ => {}
        }

        // Show extracted hashtags before processing
        let tags = hashtags::extract_hashtags(input);
        if !tags.is_empty() {
            println!("  🏷️  {}", tags.join(" "));
        }

        match engine.process_query(input) {
            Ok(response) => {
                let expert = engine.active_expert();
                let icon = match expert {
                    "structural" => "🏗️",
                    "flow_error" => "🔀",
                    "system_io" => "📁",
                    "rust_coding" => "🦀",
                    "friendly_chat" => "💬",
                    "teaching" => "📚",
                    _ => "🤖",
                };
                println!("\n[{icon} {expert}]");
                println!("{}\n", response);
            }
            Err(e) => {
                eprintln!("Error: {:#}", e);
            }
        }
    }

    Ok(())
}