memvid_cli/

config.rs

//! CLI configuration and environment handling
//!
//! This module provides configuration loading from environment variables,
//! tracing initialization, and embedding runtime management for semantic search.

use std::env;
use std::path::PathBuf;
use std::str::FromStr;

use anyhow::{anyhow, Result};
use ed25519_dalek::VerifyingKey;

const DEFAULT_API_URL: &str = "https://kgpfm35ddc.us-east-2.awsapprunner.com";
const DEFAULT_CACHE_DIR: &str = "~/.cache/memvid";

/// Supported embedding models for semantic search
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum EmbeddingModelChoice {
    /// BGE-small-en-v1.5: Fast, 384-dim, ~78% accuracy (default)
    #[default]
    BgeSmall,
    /// BGE-base-en-v1.5: Balanced, 768-dim, ~85% accuracy
    BgeBase,
    /// Nomic-embed-text-v1.5: High accuracy, 768-dim, ~86% accuracy
    Nomic,
    /// GTE-large-en-v1.5: Best semantic depth, 1024-dim
    GteLarge,
    /// OpenAI text-embedding-3-large: Highest quality, 3072-dim (requires OPENAI_API_KEY)
    OpenAILarge,
    /// OpenAI text-embedding-3-small: Good quality, 1536-dim (requires OPENAI_API_KEY)
    OpenAISmall,
    /// OpenAI text-embedding-ada-002: Legacy model, 1536-dim (requires OPENAI_API_KEY)
    OpenAIAda,
}

impl EmbeddingModelChoice {
    /// Check if this is an OpenAI model (requires OPENAI_API_KEY)
    pub fn is_openai(&self) -> bool {
        matches!(
            self,
            EmbeddingModelChoice::OpenAILarge
                | EmbeddingModelChoice::OpenAISmall
                | EmbeddingModelChoice::OpenAIAda
        )
    }

    /// Get the fastembed EmbeddingModel enum value (only for local models)
    ///
    /// # Panics
    /// Panics if called on an OpenAI model. Use `is_openai()` to check first.
    pub fn to_fastembed_model(&self) -> fastembed::EmbeddingModel {
        match self {
            EmbeddingModelChoice::BgeSmall => fastembed::EmbeddingModel::BGESmallENV15,
            EmbeddingModelChoice::BgeBase => fastembed::EmbeddingModel::BGEBaseENV15,
            EmbeddingModelChoice::Nomic => fastembed::EmbeddingModel::NomicEmbedTextV15,
            EmbeddingModelChoice::GteLarge => fastembed::EmbeddingModel::GTELargeENV15,
            EmbeddingModelChoice::OpenAILarge
            | EmbeddingModelChoice::OpenAISmall
            | EmbeddingModelChoice::OpenAIAda => {
                panic!("OpenAI models don't use fastembed. Check is_openai() first.")
            }
        }
    }

    /// Get human-readable model name
    pub fn name(&self) -> &'static str {
        match self {
            EmbeddingModelChoice::BgeSmall => "bge-small",
            EmbeddingModelChoice::BgeBase => "bge-base",
            EmbeddingModelChoice::Nomic => "nomic",
            EmbeddingModelChoice::GteLarge => "gte-large",
            EmbeddingModelChoice::OpenAILarge => "openai-large",
            EmbeddingModelChoice::OpenAISmall => "openai-small",
            EmbeddingModelChoice::OpenAIAda => "openai-ada",
        }
    }

    /// Get embedding dimensions
    pub fn dimensions(&self) -> usize {
        match self {
            EmbeddingModelChoice::BgeSmall => 384,
            EmbeddingModelChoice::BgeBase => 768,
            EmbeddingModelChoice::Nomic => 768,
            EmbeddingModelChoice::GteLarge => 1024,
            EmbeddingModelChoice::OpenAILarge => 3072,
            EmbeddingModelChoice::OpenAISmall => 1536,
            EmbeddingModelChoice::OpenAIAda => 1536,
        }
    }
}

impl FromStr for EmbeddingModelChoice {
    type Err = anyhow::Error;

    fn from_str(s: &str) -> Result<Self> {
        match s.to_lowercase().as_str() {
            "bge-small" | "bge_small" | "bgesmall" | "small" => Ok(EmbeddingModelChoice::BgeSmall),
            "bge-base" | "bge_base" | "bgebase" | "base" => Ok(EmbeddingModelChoice::BgeBase),
            "nomic" | "nomic-embed" | "nomic_embed" => Ok(EmbeddingModelChoice::Nomic),
            "gte-large" | "gte_large" | "gtelarge" | "gte" => Ok(EmbeddingModelChoice::GteLarge),
            // OpenAI models - default "openai" maps to "openai-large" for highest quality
            "openai" | "openai-large" | "openai_large" | "text-embedding-3-large" => {
                Ok(EmbeddingModelChoice::OpenAILarge)
            }
            "openai-small" | "openai_small" | "text-embedding-3-small" => {
                Ok(EmbeddingModelChoice::OpenAISmall)
            }
            "openai-ada" | "openai_ada" | "text-embedding-ada-002" | "ada" => {
                Ok(EmbeddingModelChoice::OpenAIAda)
            }
            _ => Err(anyhow!(
                "unknown embedding model '{}'. Valid options: bge-small, bge-base, nomic, gte-large, openai, openai-small, openai-ada",
                s
            )),
        }
    }
}

impl std::fmt::Display for EmbeddingModelChoice {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "{}", self.name())
    }
}

impl EmbeddingModelChoice {
    /// Infer the best embedding model from vector dimension stored in MV2 file.
    ///
    /// This enables auto-detection: users don't need to specify --query-embedding-model
    /// if the MV2 file has vectors. The dimension uniquely identifies the model family.
    ///
    /// # Dimension Mapping
    /// - 384  → BGE-small (default local model)
    /// - 768  → BGE-base (could also be Nomic, but same dimension works)
    /// - 1024 → GTE-large
    /// - 1536 → OpenAI small/ada
    /// - 3072 → OpenAI large
    pub fn from_dimension(dim: u32) -> Option<Self> {
        match dim {
            384 => Some(EmbeddingModelChoice::BgeSmall),
            768 => Some(EmbeddingModelChoice::BgeBase), // Could be Nomic, but same dim
            1024 => Some(EmbeddingModelChoice::GteLarge),
            1536 => Some(EmbeddingModelChoice::OpenAISmall), // Could be Ada, same dim
            3072 => Some(EmbeddingModelChoice::OpenAILarge),
            0 => None, // No vectors in file
            _ => {
                tracing::warn!(
                    "Unknown embedding dimension {}, using default model",
                    dim
                );
                None
            }
        }
    }
}

/// CLI configuration loaded from environment variables
#[derive(Debug, Clone)]
pub struct CliConfig {
    pub api_key: Option<String>,
    pub api_url: String,
    pub cache_dir: PathBuf,
    pub ticket_pubkey: Option<VerifyingKey>,
    pub models_dir: PathBuf,
    pub offline: bool,
    /// Embedding model for semantic search (can be overridden by CLI flag)
    pub embedding_model: EmbeddingModelChoice,
}

impl PartialEq for CliConfig {
    fn eq(&self, other: &Self) -> bool {
        self.api_key == other.api_key
            && self.api_url == other.api_url
            && self.cache_dir == other.cache_dir
            && self.models_dir == other.models_dir
            && self.offline == other.offline
            && self.embedding_model == other.embedding_model
    }
}

impl Eq for CliConfig {}

impl CliConfig {
    pub fn load() -> Result<Self> {
        let api_key = env::var("MEMVID_API_KEY").ok().and_then(|value| {
            let trimmed = value.trim().to_string();
            (!trimmed.is_empty()).then_some(trimmed)
        });

        let api_url = env::var("MEMVID_API_URL").unwrap_or_else(|_| DEFAULT_API_URL.to_string());

        let cache_dir_raw =
            env::var("MEMVID_CACHE_DIR").unwrap_or_else(|_| DEFAULT_CACHE_DIR.to_string());
        let cache_dir = expand_path(&cache_dir_raw)?;

        let models_dir_raw =
            env::var("MEMVID_MODELS_DIR").unwrap_or_else(|_| "~/.memvid/models".to_string());
        let models_dir = expand_path(&models_dir_raw)?;

        let ticket_pubkey = env::var("MEMVID_TICKET_PUBKEY")
            .ok()
            .and_then(|value| {
                let trimmed = value.trim();
                if trimmed.is_empty() {
                    None
                } else {
                    Some(memvid_core::parse_ed25519_public_key_base64(trimmed))
                }
            })
            .transpose()?;

        let offline = env::var("MEMVID_OFFLINE")
            .ok()
            .map(|value| match value.trim().to_ascii_lowercase().as_str() {
                "1" | "true" | "yes" => true,
                _ => false,
            })
            .unwrap_or(false);

        // Load embedding model from env var, default to BGE-small
        let embedding_model = env::var("MEMVID_EMBEDDING_MODEL")
            .ok()
            .and_then(|value| {
                let trimmed = value.trim();
                if trimmed.is_empty() {
                    None
                } else {
                    EmbeddingModelChoice::from_str(trimmed).ok()
                }
            })
            .unwrap_or_default();

        Ok(Self {
            api_key,
            api_url,
            cache_dir,
            ticket_pubkey,
            models_dir,
            offline,
            embedding_model,
        })
    }

    /// Create a new config with a different embedding model
    pub fn with_embedding_model(&self, model: EmbeddingModelChoice) -> Self {
        Self {
            embedding_model: model,
            ..self.clone()
        }
    }
}

fn expand_path(value: &str) -> Result<PathBuf> {
    if value.trim().is_empty() {
        return Err(anyhow!("cache directory cannot be empty"));
    }

    let expanded = if let Some(stripped) = value.strip_prefix("~/") {
        home_dir()?.join(stripped)
    } else if let Some(stripped) = value.strip_prefix("~\\") {
        // Support Windows-style "~\" prefix.
        home_dir()?.join(stripped)
    } else if value == "~" {
        home_dir()?
    } else {
        PathBuf::from(value)
    };

    if expanded.is_absolute() {
        Ok(expanded)
    } else {
        Ok(env::current_dir()?.join(expanded))
    }
}

fn home_dir() -> Result<PathBuf> {
    if let Some(path) = env::var_os("HOME") {
        if !path.is_empty() {
            return Ok(PathBuf::from(path));
        }
    }

    #[cfg(windows)]
    {
        if let Some(path) = env::var_os("USERPROFILE") {
            if !path.is_empty() {
                return Ok(PathBuf::from(path));
            }
        }
        if let (Some(drive), Some(path)) = (env::var_os("HOMEDRIVE"), env::var_os("HOMEPATH")) {
            if !drive.is_empty() && !path.is_empty() {
                return Ok(PathBuf::from(format!(
                    "{}{}",
                    drive.to_string_lossy(),
                    path.to_string_lossy()
                )));
            }
        }
    }

    Err(anyhow!("unable to resolve home directory"))
}

#[cfg(test)]
mod tests {
    use super::*;
    use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;
    use base64::Engine;
    use ed25519_dalek::SigningKey;
    use std::sync::{Mutex, OnceLock};

    fn env_lock() -> std::sync::MutexGuard<'static, ()> {
        static LOCK: OnceLock<Mutex<()>> = OnceLock::new();
        LOCK.get_or_init(|| Mutex::new(())).lock().unwrap()
    }

    fn set_or_unset(var: &str, value: Option<String>) {
        match value {
            Some(v) => unsafe { env::set_var(var, v) },
            None => unsafe { env::remove_var(var) },
        }
    }

    #[test]
    fn defaults_expand_using_home_directory() {
        let _guard = env_lock();

        let previous_home = env::var("HOME").ok();
        #[cfg(windows)]
        let previous_userprofile = env::var("USERPROFILE").ok();

        for var in [
            "MEMVID_API_KEY",
            "MEMVID_API_URL",
            "MEMVID_CACHE_DIR",
            "MEMVID_TICKET_PUBKEY",
            "MEMVID_MODELS_DIR",
            "MEMVID_OFFLINE",
        ] {
            unsafe { env::remove_var(var) };
        }

        let tmp = tempfile::tempdir().expect("tmpdir");
        let tmp_path = tmp.path().to_path_buf();
        unsafe { env::set_var("HOME", &tmp_path) };
        #[cfg(windows)]
        unsafe {
            env::set_var("USERPROFILE", &tmp_path)
        };

        let config = CliConfig::load().expect("load");
        assert_eq!(config.api_key, None);
        assert_eq!(config.api_url, DEFAULT_API_URL);
        assert_eq!(config.cache_dir, tmp_path.join(".cache/memvid"));
        assert!(config.ticket_pubkey.is_none());
        assert_eq!(config.models_dir, tmp_path.join(".memvid/models"));
        assert!(!config.offline);

        set_or_unset("HOME", previous_home);
        #[cfg(windows)]
        {
            set_or_unset("USERPROFILE", previous_userprofile);
        }
    }

    #[test]
    fn env_overrides_are_respected() {
        let _guard = env_lock();

        let previous_env: Vec<(&'static str, Option<String>)> = [
            "MEMVID_API_KEY",
            "MEMVID_API_URL",
            "MEMVID_CACHE_DIR",
            "MEMVID_TICKET_PUBKEY",
            "MEMVID_MODELS_DIR",
            "MEMVID_OFFLINE",
        ]
        .into_iter()
        .map(|var| (var, env::var(var).ok()))
        .collect();

        unsafe { env::set_var("MEMVID_API_KEY", "abc123") };
        unsafe { env::set_var("MEMVID_API_URL", "https://staging.memvid.app") };
        unsafe { env::set_var("MEMVID_CACHE_DIR", "~/memvid-cache") };
        unsafe { env::set_var("MEMVID_MODELS_DIR", "~/models") };
        unsafe { env::set_var("MEMVID_OFFLINE", "true") };
        let signing = SigningKey::from_bytes(&[9u8; 32]);
        let encoded = BASE64_STANDARD.encode(signing.verifying_key().as_bytes());
        unsafe { env::set_var("MEMVID_TICKET_PUBKEY", encoded) };

        let tmp = tempfile::tempdir().expect("tmpdir");
        let tmp_path = tmp.path().to_path_buf();
        unsafe { env::set_var("HOME", &tmp_path) };
        #[cfg(windows)]
        unsafe {
            env::set_var("USERPROFILE", &tmp_path)
        };

        let config = CliConfig::load().expect("load");
        assert_eq!(config.api_key.as_deref(), Some("abc123"));
        assert_eq!(config.api_url, "https://staging.memvid.app");
        assert_eq!(config.cache_dir, tmp_path.join("memvid-cache"));
        assert_eq!(
            config.ticket_pubkey.expect("pubkey").as_bytes(),
            signing.verifying_key().as_bytes()
        );
        assert_eq!(config.models_dir, tmp_path.join("models"));
        assert!(config.offline);

        for (var, value) in previous_env {
            set_or_unset(var, value);
        }
    }

    #[test]
    fn rejects_empty_cache_dir() {
        let _guard = env_lock();

        let previous = env::var("MEMVID_CACHE_DIR").ok();
        unsafe { env::set_var("MEMVID_CACHE_DIR", " ") };
        let err = CliConfig::load().expect_err("should fail");
        assert!(err.to_string().contains("cache directory"));
        set_or_unset("MEMVID_CACHE_DIR", previous);
    }
}

/// Initialize tracing/logging based on verbosity level
pub fn init_tracing(verbosity: u8) -> Result<()> {
    use std::io::IsTerminal;
    use tracing_subscriber::{filter::Directive, fmt, EnvFilter};

    let level = match verbosity {
        0 => "warn",
        1 => "info",
        2 => "debug",
        _ => "trace",
    };

    let mut env_filter =
        EnvFilter::try_from_default_env().unwrap_or_else(|_| EnvFilter::new(level));
    for directive_str in ["llama_cpp=error", "llama_cpp_sys=error", "ggml=error"] {
        if let Ok(directive) = directive_str.parse::<Directive>() {
            env_filter = env_filter.add_directive(directive);
        }
    }

    // Disable ANSI color codes when stderr is not a terminal (e.g., piped or
    // combined with `2>&1`). This prevents control characters from polluting
    // JSON output when combined with stdout.
    let use_ansi = std::io::stderr().is_terminal();

    fmt()
        .with_env_filter(env_filter)
        .with_writer(std::io::stderr)
        .with_target(false)
        .without_time()
        .with_ansi(use_ansi)
        .try_init()
        .map_err(|err| anyhow!(err))?;
    Ok(())
}

/// Resolve LLM context budget override from CLI or environment
pub fn resolve_llm_context_budget_override(cli_value: Option<usize>) -> Result<Option<usize>> {
    use anyhow::bail;

    if let Some(value) = cli_value {
        if value == 0 {
            bail!("--llm-context-depth must be a positive integer");
        }
        return Ok(Some(value));
    }

    let raw_env = match env::var("MEMVID_LLM_CONTEXT_BUDGET") {
        Ok(value) => value,
        Err(_) => return Ok(None),
    };

    let trimmed = raw_env.trim();
    if trimmed.is_empty() {
        return Ok(None);
    }

    let digits: String = trimmed
        .chars()
        .filter(|ch| !ch.is_ascii_whitespace() && *ch != '_')
        .collect();

    if digits.is_empty() {
        bail!("MEMVID_LLM_CONTEXT_BUDGET must be a positive integer value");
    }

    let value: usize = digits.parse().map_err(|err| {
        anyhow!(
            "MEMVID_LLM_CONTEXT_BUDGET value '{}' is not a valid number: {}",
            trimmed,
            err
        )
    })?;

    if value == 0 {
        bail!("MEMVID_LLM_CONTEXT_BUDGET must be a positive integer");
    }

    Ok(Some(value))
}

use crate::openai_embeddings::OpenAIEmbeddingProvider;

/// Internal embedding backend - either local fastembed or OpenAI API
#[derive(Clone)]
enum EmbeddingBackend {
    FastEmbed(std::sync::Arc<std::sync::Mutex<fastembed::TextEmbedding>>),
    OpenAI(std::sync::Arc<OpenAIEmbeddingProvider>),
}

/// Embedding runtime wrapper supporting both local and OpenAI embeddings
#[derive(Clone)]
pub struct EmbeddingRuntime {
    backend: EmbeddingBackend,
    dimension: usize,
}

impl EmbeddingRuntime {
    fn new_fastembed(model: fastembed::TextEmbedding, dimension: usize) -> Self {
        Self {
            backend: EmbeddingBackend::FastEmbed(std::sync::Arc::new(std::sync::Mutex::new(
                model,
            ))),
            dimension,
        }
    }

    fn new_openai(provider: OpenAIEmbeddingProvider, dimension: usize) -> Self {
        Self {
            backend: EmbeddingBackend::OpenAI(std::sync::Arc::new(provider)),
            dimension,
        }
    }

    /// Maximum characters for embedding text to avoid exceeding OpenAI's 8192 token limit.
    /// Using ~3 chars/token estimate (conservative for dense content), 20K chars ≈ 6.6K tokens.
    const MAX_EMBEDDING_TEXT_LEN: usize = 20_000;

    /// Truncate text for embedding to fit within token limits.
    fn truncate_for_embedding(text: &str) -> std::borrow::Cow<'_, str> {
        if text.len() <= Self::MAX_EMBEDDING_TEXT_LEN {
            std::borrow::Cow::Borrowed(text)
        } else {
            // Find the last valid UTF-8 char boundary within the limit
            let truncated = &text[..Self::MAX_EMBEDDING_TEXT_LEN];
            let end = truncated
                .char_indices()
                .rev()
                .next()
                .map(|(i, c)| i + c.len_utf8())
                .unwrap_or(Self::MAX_EMBEDDING_TEXT_LEN);
            tracing::info!(
                "Truncated embedding text from {} to {} chars to fit OpenAI token limit",
                text.len(),
                end
            );
            std::borrow::Cow::Owned(text[..end].to_string())
        }
    }

    pub fn embed(&self, text: &str) -> Result<Vec<f32>> {
        // Truncate text for OpenAI to avoid token limit errors
        let text = match &self.backend {
            EmbeddingBackend::OpenAI(_) => Self::truncate_for_embedding(text),
            _ => std::borrow::Cow::Borrowed(text),
        };

        match &self.backend {
            EmbeddingBackend::FastEmbed(model) => {
                let mut guard = model
                    .lock()
                    .map_err(|_| anyhow!("fastembed runtime poisoned"))?;
                let outputs = guard
                    .embed(vec![text.into_owned()], None)
                    .map_err(|err| anyhow!("failed to compute embedding with fastembed: {err}"))?;
                outputs
                    .into_iter()
                    .next()
                    .ok_or_else(|| anyhow!("fastembed returned no embedding output"))
            }
            EmbeddingBackend::OpenAI(provider) => {
                use memvid_core::EmbeddingProvider;
                provider
                    .embed_text(&text)
                    .map_err(|err| anyhow!("failed to compute embedding with OpenAI: {err}"))
            }
        }
    }

    pub fn dimension(&self) -> usize {
        self.dimension
    }
}

impl memvid_core::VecEmbedder for EmbeddingRuntime {
    fn embed_query(&self, text: &str) -> memvid_core::Result<Vec<f32>> {
        self.embed(text)
            .map_err(|err| memvid_core::MemvidError::EmbeddingFailed {
                reason: err.to_string().into_boxed_str(),
            })
    }

    fn embedding_dimension(&self) -> usize {
        self.dimension()
    }
}

/// Ensure fastembed cache directory exists
fn ensure_fastembed_cache(config: &CliConfig) -> Result<PathBuf> {
    use std::fs;

    let cache_dir = config.models_dir.clone();
    fs::create_dir_all(&cache_dir)?;
    Ok(cache_dir)
}

/// Get approximate model size in MB for user-friendly error messages
fn model_size_mb(model: EmbeddingModelChoice) -> usize {
    match model {
        EmbeddingModelChoice::BgeSmall => 33,
        EmbeddingModelChoice::BgeBase => 110,
        EmbeddingModelChoice::Nomic => 137,
        EmbeddingModelChoice::GteLarge => 327,
        // OpenAI models don't require local download
        EmbeddingModelChoice::OpenAILarge
        | EmbeddingModelChoice::OpenAISmall
        | EmbeddingModelChoice::OpenAIAda => 0,
    }
}

/// Instantiate an embedding runtime with the configured model
fn instantiate_embedding_runtime(config: &CliConfig) -> Result<EmbeddingRuntime> {
    use tracing::info;

    let embedding_model = config.embedding_model;

    info!(
        "Loading embedding model: {} ({}D)",
        embedding_model.name(),
        embedding_model.dimensions()
    );

    // Check if OpenAI model
    if embedding_model.is_openai() {
        return instantiate_openai_runtime(embedding_model);
    }

    // Local fastembed model
    instantiate_fastembed_runtime(config, embedding_model)
}

/// Instantiate OpenAI embedding runtime
fn instantiate_openai_runtime(embedding_model: EmbeddingModelChoice) -> Result<EmbeddingRuntime> {
    use anyhow::bail;
    use memvid_core::EmbeddingConfig;
    use tracing::info;

    let api_key = std::env::var("OPENAI_API_KEY")
        .map_err(|_| anyhow!("OPENAI_API_KEY environment variable is required for OpenAI embeddings"))?;

    if api_key.is_empty() {
        bail!("OPENAI_API_KEY cannot be empty");
    }

    let config = match embedding_model {
        EmbeddingModelChoice::OpenAILarge => EmbeddingConfig::openai_large(),
        EmbeddingModelChoice::OpenAISmall => EmbeddingConfig::openai_small(),
        EmbeddingModelChoice::OpenAIAda => EmbeddingConfig::openai_ada(),
        _ => unreachable!("is_openai() should have been false"),
    };

    let provider = OpenAIEmbeddingProvider::new(api_key, config.clone())
        .map_err(|err| anyhow!("failed to create OpenAI embedding provider: {err}"))?;

    info!(
        "OpenAI embedding provider ready: model={}, dimension={}",
        config.model, config.dimension
    );

    Ok(EmbeddingRuntime::new_openai(provider, config.dimension))
}

/// Instantiate fastembed (local) embedding runtime
fn instantiate_fastembed_runtime(
    config: &CliConfig,
    embedding_model: EmbeddingModelChoice,
) -> Result<EmbeddingRuntime> {
    use anyhow::bail;
    use fastembed::{InitOptions, TextEmbedding};
    use std::fs;

    let cache_dir = ensure_fastembed_cache(config)?;

    if config.offline {
        let mut entries = fs::read_dir(&cache_dir)?;
        if entries.next().is_none() {
            bail!(
                "semantic embeddings unavailable while offline; allow one connected run so fastembed can cache model weights"
            );
        }
    }

    let options = InitOptions::new(embedding_model.to_fastembed_model())
        .with_cache_dir(cache_dir)
        .with_show_download_progress(true);
    let mut model = TextEmbedding::try_new(options).map_err(|err| {
        // Provide platform-specific guidance for model download issues
        let platform_hint = if cfg!(target_os = "windows") {
            "\n\nWindows users: If model downloads fail, try:\n\
            1. Run as Administrator\n\
            2. Check your antivirus isn't blocking downloads\n\
            3. Use OpenAI embeddings instead: set OPENAI_API_KEY and use --embedding-model openai"
        } else if cfg!(target_os = "linux") {
            "\n\nLinux users: If model downloads fail, try:\n\
            1. Check disk space in ~/.memvid/models\n\
            2. Ensure you have network access to huggingface.co\n\
            3. Use OpenAI embeddings instead: export OPENAI_API_KEY=... and use --embedding-model openai"
        } else {
            "\n\nIf model downloads fail, try using OpenAI embeddings:\n\
            export OPENAI_API_KEY=your-key && memvid ... --embedding-model openai"
        };

        anyhow!(
            "Failed to initialize embedding model '{}': {err}\n\n\
            This typically means the model couldn't be downloaded or loaded.\n\
            Model size: ~{} MB{}\n\n\
            See: https://docs.memvid.com/embedding-models",
            embedding_model.name(),
            model_size_mb(embedding_model),
            platform_hint
        )
    })?;

    let probe = model
        .embed(vec!["memvid probe".to_string()], None)
        .map_err(|err| anyhow!("failed to determine embedding dimension: {err}"))?;
    let dimension = probe.first().map(|vec| vec.len()).unwrap_or(0);

    if dimension == 0 {
        bail!("fastembed reported zero-length embeddings");
    }

    // Verify dimension matches expected
    if dimension != embedding_model.dimensions() {
        tracing::warn!(
            "Embedding dimension mismatch: expected {}, got {}",
            embedding_model.dimensions(),
            dimension
        );
    }

    Ok(EmbeddingRuntime::new_fastembed(model, dimension))
}

/// Load embedding runtime (fails if unavailable)
pub fn load_embedding_runtime(config: &CliConfig) -> Result<EmbeddingRuntime> {
    use anyhow::bail;

    match instantiate_embedding_runtime(config) {
        Ok(runtime) => Ok(runtime),
        Err(err) => {
            if config.offline {
                bail!(
                    "semantic embeddings unavailable while offline; allow one connected run so fastembed can cache model weights ({err})"
                );
            }
            Err(err)
        }
    }
}

/// Try to load embedding runtime (returns None if unavailable)
pub fn try_load_embedding_runtime(config: &CliConfig) -> Option<EmbeddingRuntime> {
    use tracing::warn;

    match instantiate_embedding_runtime(config) {
        Ok(runtime) => Some(runtime),
        Err(err) => {
            warn!("semantic embeddings unavailable: {err}");
            None
        }
    }
}

/// Load embedding runtime with an optional model override.
/// If `model_override` is provided, it will be used instead of the config's embedding_model.
pub fn load_embedding_runtime_with_model(
    config: &CliConfig,
    model_override: Option<&str>,
) -> Result<EmbeddingRuntime> {
    use tracing::info;

    let embedding_model = match model_override {
        Some(model_str) => {
            let parsed = model_str.parse::<EmbeddingModelChoice>()?;
            info!("Using embedding model override: {} ({}D)", parsed.name(), parsed.dimensions());
            parsed
        }
        None => config.embedding_model,
    };

    info!(
        "Loading embedding model: {} ({}D)",
        embedding_model.name(),
        embedding_model.dimensions()
    );

    if embedding_model.is_openai() {
        return instantiate_openai_runtime(embedding_model);
    }

    instantiate_fastembed_runtime(config, embedding_model)
}

/// Try to load embedding runtime with model override (returns None if unavailable)
pub fn try_load_embedding_runtime_with_model(
    config: &CliConfig,
    model_override: Option<&str>,
) -> Option<EmbeddingRuntime> {
    use tracing::warn;

    match load_embedding_runtime_with_model(config, model_override) {
        Ok(runtime) => Some(runtime),
        Err(err) => {
            warn!("semantic embeddings unavailable: {err}");
            None
        }
    }
}

/// Load embedding runtime by auto-detecting from MV2 vector dimension.
///
/// Priority:
/// 1. Explicit model override (--query-embedding-model flag)
/// 2. Auto-detect from MV2 file's stored dimension
/// 3. Fall back to config default
///
/// This allows users to omit --query-embedding-model when querying files
/// created with non-default embedding models (like OpenAI).
pub fn load_embedding_runtime_for_mv2(
    config: &CliConfig,
    model_override: Option<&str>,
    mv2_dimension: Option<u32>,
) -> Result<EmbeddingRuntime> {
    use tracing::info;

    // Priority 1: Explicit override
    if let Some(model_str) = model_override {
        return load_embedding_runtime_with_model(config, Some(model_str));
    }

    // Priority 2: Auto-detect from MV2 dimension
    if let Some(dim) = mv2_dimension {
        if let Some(detected_model) = EmbeddingModelChoice::from_dimension(dim) {
            info!(
                "Auto-detected embedding model from MV2: {} ({}D)",
                detected_model.name(),
                dim
            );

            // For OpenAI models, check if API key is available
            if detected_model.is_openai() {
                if std::env::var("OPENAI_API_KEY").is_ok() {
                    return load_embedding_runtime_with_model(config, Some(detected_model.name()));
                } else {
                    // OpenAI detected but no API key - provide helpful error
                    return Err(anyhow!(
                        "MV2 file uses OpenAI embeddings ({}D) but OPENAI_API_KEY is not set.\n\n\
                        Options:\n\
                        1. Set OPENAI_API_KEY environment variable\n\
                        2. Use --query-embedding-model to specify a different model\n\
                        3. Use lexical-only search with --mode lex\n\n\
                        See: https://docs.memvid.com/embedding-models",
                        dim
                    ));
                }
            }

            return load_embedding_runtime_with_model(config, Some(detected_model.name()));
        }
    }

    // Priority 3: Fall back to config default
    load_embedding_runtime(config)
}

/// Try to load embedding runtime for MV2 with auto-detection (returns None if unavailable)
pub fn try_load_embedding_runtime_for_mv2(
    config: &CliConfig,
    model_override: Option<&str>,
    mv2_dimension: Option<u32>,
) -> Option<EmbeddingRuntime> {
    use tracing::warn;

    match load_embedding_runtime_for_mv2(config, model_override, mv2_dimension) {
        Ok(runtime) => Some(runtime),
        Err(err) => {
            warn!("semantic embeddings unavailable: {err}");
            None
        }
    }
}