engram-memory 0.2.0

use std::fs;
use std::path::{Path, PathBuf};
use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};

use serde::Deserialize;

use engram_hnsw::HnswParams;
use engram_llm_client::{
    EmbeddingProvider, LocalTextGenerator, OpenAITextGenerator, TextGenerator,
    VoyageEmbeddingProvider,
};

use crate::error::CoreError;

const CONFIG_HOME_SUBDIR: &str = ".engram/engram.toml";
const PROJECT_DIR_MARKER: &str = ".engram";

const DEFAULT_DB_PATH: &str = "~/.engram/memories.db";
const DEFAULT_SOCKET_PATH: &str = "~/.engram/engram.sock";
const DEFAULT_EMBEDDING_PROVIDER: &str = "voyage";
const DEFAULT_EMBEDDING_MODEL: &str = "voyage-code-3";
const DEFAULT_EMBEDDING_DIMENSION: usize = 1024;
const DEFAULT_HYDE_THRESHOLD: usize = 0;
const DEFAULT_LLM_PROVIDER: &str = "openai";
const DEFAULT_LLM_MODEL: &str = "gpt-4o-mini";
const DEFAULT_REINDEX_INTERVAL_SECS: u64 = 3600;
const DEFAULT_HNSW_MAX_CONNECTIONS: usize = 16;
const DEFAULT_HNSW_EF_CONSTRUCTION: usize = 200;
const DEFAULT_HNSW_EF_SEARCH: usize = 40;
const DEFAULT_CONSOLIDATION_STALE_DAYS: u32 = 90;
const DEFAULT_CONSOLIDATION_MIN_SCORE: f64 = 0.3;
const DEFAULT_TRAINER_BINARY: &str = "engram-trainer";
const DEFAULT_TRAINER_TIMEOUT_SECS: u64 = 300;
const DEFAULT_MODELS_PATH: &str = "~/.engram/models";

#[derive(Deserialize, Clone)]
pub struct Config {
    pub database: DatabaseConfig,
    pub embedding: EmbeddingConfig,
    pub llm: LlmConfig,
    pub server: ServerConfig,
    pub hnsw: HnswConfig,
    #[serde(default)]
    pub consolidation: ConsolidationConfig,
    #[serde(default)]
    pub trainer: TrainerConfig,
}

#[derive(Deserialize, Clone)]
pub struct DatabaseConfig {
    pub path: String,
}

#[derive(Deserialize, Clone)]
pub struct EmbeddingConfig {
    pub provider: String,
    pub api_key: Option<String>,
    pub model: Option<String>,
    pub dimension: Option<usize>,
    #[serde(default)]
    pub hyde_threshold: usize,
}

#[derive(Deserialize, Clone)]
pub struct LlmConfig {
    pub provider: String,
    pub api_key: Option<String>,
    pub model: Option<String>,
}

#[derive(Deserialize, Clone)]
pub struct ServerConfig {
    pub socket_path: String,
    pub reindex_interval_secs: u64,
}

#[derive(Deserialize, Clone)]
pub struct HnswConfig {
    pub max_connections: usize,
    pub ef_construction: usize,
    pub ef_search: usize,
    pub dimension: usize,
}

#[derive(Deserialize, Clone)]
pub struct ConsolidationConfig {
    pub stale_days: u32,
    pub min_score: f64,
}

#[derive(Deserialize, Clone)]
pub struct TrainerConfig {
    pub trainer_binary: String,
    pub trainer_timeout_secs: u64,
    pub models_path: String,
}

impl Default for TrainerConfig {
    fn default() -> Self {
        Self {
            trainer_binary: DEFAULT_TRAINER_BINARY.into(),
            trainer_timeout_secs: DEFAULT_TRAINER_TIMEOUT_SECS,
            models_path: DEFAULT_MODELS_PATH.into(),
        }
    }
}

impl Default for ConsolidationConfig {
    fn default() -> Self {
        Self {
            stale_days: DEFAULT_CONSOLIDATION_STALE_DAYS,
            min_score: DEFAULT_CONSOLIDATION_MIN_SCORE,
        }
    }
}

impl Config {
    pub fn load() -> Result<Self, CoreError> {
        if let Some(home) = home_directory() {
            let home_config = Path::new(&home).join(CONFIG_HOME_SUBDIR);
            if home_config.exists() {
                return Self::load_from_path(home_config.to_str().unwrap_or_default());
            }
        }

        let mut config = Self::default();
        config.apply_env_overrides();
        Ok(config)
    }

    pub fn load_from_path(path: &str) -> Result<Self, CoreError> {
        let content = fs::read_to_string(path).map_err(|_| CoreError::ConfigNotFound)?;
        let mut config: Config = toml::from_str(&content)
            .map_err(|error| CoreError::ConfigParseError(error.to_string()))?;
        config.apply_env_overrides();
        Ok(config)
    }

    pub fn build_embedding_provider(
        &self,
    ) -> Result<Box<dyn EmbeddingProvider + Send + Sync>, CoreError> {
        let dimension = self
            .embedding
            .dimension
            .unwrap_or(DEFAULT_EMBEDDING_DIMENSION);
        match self.embedding.provider.as_str() {
            "voyage" => {
                let api_key = self
                    .embedding
                    .api_key
                    .as_deref()
                    .filter(|key| !key.is_empty())
                    .ok_or_else(|| CoreError::InvalidProvider("voyage requires api_key".into()))?;
                let provider = VoyageEmbeddingProvider::new(api_key.to_owned())?;
                Ok(Box::new(provider))
            }
            "deterministic" => Ok(Box::new(DeterministicEmbeddingProvider { dimension })),
            other => Err(CoreError::InvalidProvider(format!(
                "{other} embedding not supported"
            ))),
        }
    }

    pub fn build_text_generator(&self) -> Result<Box<dyn TextGenerator + Send + Sync>, CoreError> {
        match self.llm.provider.as_str() {
            "openai" => {
                let api_key = self
                    .llm
                    .api_key
                    .as_deref()
                    .filter(|key| !key.is_empty())
                    .ok_or_else(|| CoreError::InvalidProvider("openai requires api_key".into()))?;
                let generator = OpenAITextGenerator::new(api_key.to_owned())?;
                Ok(Box::new(generator))
            }
            "local" => {
                let models_path = expand_tilde(&self.trainer.models_path);
                let model_path = format!("{models_path}/text_generator.onnx");
                let tokenizer_path = format!("{models_path}/tokenizer.json");
                let generator = LocalTextGenerator::new(&model_path, &tokenizer_path)
                    .map_err(|error| CoreError::InvalidProvider(error.to_string()))?;
                Ok(Box::new(generator))
            }
            other => Err(CoreError::InvalidProvider(format!(
                "{other} llm not supported"
            ))),
        }
    }

    pub fn build_hnsw_params(&self) -> Result<HnswParams, CoreError> {
        HnswParams::new(self.hnsw.dimension)?
            .with_max_connections(self.hnsw.max_connections)?
            .with_ef_construction(self.hnsw.ef_construction)?
            .with_ef_search(self.hnsw.ef_search)
            .map_err(CoreError::Hnsw)
    }

    fn apply_env_overrides(&mut self) {
        // ENGRAM_DB_PATH / ENGRAM_SOCKET_PATH are honored directly by server.rs
        // (resolve_database_path / resolve_socket_path). Copying them onto
        // `self.database.path` / `self.server.socket_path` here would be redundant
        // and could mask the per-project layout derived from `project_dir`.
        if let Ok(value) = std::env::var("ENGRAM_EMBEDDING_MODEL") {
            self.embedding.model = Some(value);
        }
        if let Ok(value) = std::env::var("ENGRAM_LLM_MODEL") {
            self.llm.model = Some(value);
        }
        if let Ok(value) = std::env::var("ENGRAM_TRAINER_BINARY") {
            self.trainer.trainer_binary = value;
        }
        if let Ok(value) = std::env::var("ENGRAM_TRAINER_TIMEOUT")
            && let Ok(secs) = value.parse::<u64>()
        {
            self.trainer.trainer_timeout_secs = secs;
        }
        if let Ok(value) = std::env::var("ENGRAM_MODELS_PATH") {
            self.trainer.models_path = value;
        }
        self.apply_provider_api_key_overrides();
    }

    fn apply_provider_api_key_overrides(&mut self) {
        if self.embedding.provider == "voyage"
            && let Ok(value) = std::env::var("ENGRAM_VOYAGE_API_KEY")
        {
            self.embedding.api_key = Some(value);
        }
        if self.llm.provider == "openai"
            && let Ok(value) = std::env::var("ENGRAM_OPENAI_API_KEY")
        {
            self.llm.api_key = Some(value);
        }
    }
}

impl Default for Config {
    fn default() -> Self {
        Self {
            database: DatabaseConfig {
                path: DEFAULT_DB_PATH.into(),
            },
            embedding: EmbeddingConfig {
                provider: DEFAULT_EMBEDDING_PROVIDER.into(),
                api_key: None,
                model: Some(DEFAULT_EMBEDDING_MODEL.into()),
                dimension: Some(DEFAULT_EMBEDDING_DIMENSION),
                hyde_threshold: DEFAULT_HYDE_THRESHOLD,
            },
            llm: LlmConfig {
                provider: DEFAULT_LLM_PROVIDER.into(),
                api_key: None,
                model: Some(DEFAULT_LLM_MODEL.into()),
            },
            server: ServerConfig {
                socket_path: DEFAULT_SOCKET_PATH.into(),
                reindex_interval_secs: DEFAULT_REINDEX_INTERVAL_SECS,
            },
            hnsw: HnswConfig {
                max_connections: DEFAULT_HNSW_MAX_CONNECTIONS,
                ef_construction: DEFAULT_HNSW_EF_CONSTRUCTION,
                ef_search: DEFAULT_HNSW_EF_SEARCH,
                dimension: DEFAULT_EMBEDDING_DIMENSION,
            },
            consolidation: ConsolidationConfig::default(),
            trainer: TrainerConfig::default(),
        }
    }
}

pub fn resolve_project_dir(
    start: &Path,
    explicit_override: Option<&Path>,
) -> Result<PathBuf, CoreError> {
    if let Some(path) = explicit_override {
        return Ok(path.to_path_buf());
    }
    if let Ok(env_path) = std::env::var("ENGRAM_PROJECT_DIR") {
        let candidate = PathBuf::from(env_path);
        if candidate.is_absolute() {
            return Ok(candidate);
        }
    }
    let mut current = start.to_path_buf();
    loop {
        if current.join(PROJECT_DIR_MARKER).is_dir() {
            return Ok(current);
        }
        // `PathBuf::pop()` returns false at filesystem root, which terminates this loop safely.
        if !current.pop() {
            break;
        }
    }
    Err(CoreError::ProjectDirNotFound)
}

pub fn expand_tilde(path: &str) -> String {
    if let Some(rest) = path.strip_prefix("~/")
        && let Some(home) = home_directory()
    {
        return format!("{home}/{rest}");
    }
    path.to_string()
}

pub fn home_directory() -> Option<String> {
    std::env::var("HOME").ok()
}

struct DeterministicEmbeddingProvider {
    dimension: usize,
}

// Test instrumentation for the deterministic provider. Gated behind
// `DETERMINISTIC_PROVIDER_INSTRUMENTATION`, which defaults to OFF so
// production users (TUI demo mode, local setups without API keys) pay
// zero overhead. Tests call `enable_deterministic_provider_instrumentation`
// before parallel assertions and `disable_deterministic_provider_instrumentation`
// in teardown (typically via a drop-guard).
static DETERMINISTIC_PROVIDER_INSTRUMENTATION: AtomicBool = AtomicBool::new(false);
static DETERMINISTIC_PROVIDER_ENTRIES: AtomicUsize = AtomicUsize::new(0);
static DETERMINISTIC_PROVIDER_MAX_CONCURRENT: AtomicUsize = AtomicUsize::new(0);

/// Enable test-only instrumentation (counter + 20ms sleep) on the deterministic
/// embedding provider. Off in production by default; tests enable it before
/// reading concurrency counters.
pub fn enable_deterministic_provider_instrumentation() {
    DETERMINISTIC_PROVIDER_INSTRUMENTATION.store(true, Ordering::Relaxed);
}

/// Disable instrumentation. Tests should call this in teardown to avoid leaking
/// state across serially-run suites.
pub fn disable_deterministic_provider_instrumentation() {
    DETERMINISTIC_PROVIDER_INSTRUMENTATION.store(false, Ordering::Relaxed);
}

pub fn deterministic_provider_max_concurrent() -> usize {
    DETERMINISTIC_PROVIDER_MAX_CONCURRENT.load(Ordering::Relaxed)
}

pub fn reset_deterministic_provider_counters() {
    DETERMINISTIC_PROVIDER_ENTRIES.store(0, Ordering::Relaxed);
    DETERMINISTIC_PROVIDER_MAX_CONCURRENT.store(0, Ordering::Relaxed);
}

impl EmbeddingProvider for DeterministicEmbeddingProvider {
    fn embed(&self, text: &str) -> Result<Vec<f32>, engram_llm_client::ApiError> {
        // Test-only instrumentation: when enabled via
        // `enable_deterministic_provider_instrumentation`, each embed call
        // records concurrent entries and sleeps 20ms to widen the overlap
        // window for parallelism assertions. Production users never enable
        // this — a relaxed atomic load is ~1ns, so overhead is effectively
        // zero otherwise.
        let instrumented = DETERMINISTIC_PROVIDER_INSTRUMENTATION.load(Ordering::Relaxed);
        if instrumented {
            let entries = DETERMINISTIC_PROVIDER_ENTRIES.fetch_add(1, Ordering::SeqCst) + 1;
            DETERMINISTIC_PROVIDER_MAX_CONCURRENT.fetch_max(entries, Ordering::SeqCst);
            std::thread::sleep(std::time::Duration::from_millis(20));
        }

        let mut embedding = vec![0.0_f32; self.dimension];
        for (index, byte) in text.bytes().enumerate() {
            embedding[index % self.dimension] += byte as f32 / 255.0;
        }
        let norm: f32 = embedding
            .iter()
            .map(|v| v * v)
            .sum::<f32>()
            .sqrt()
            .max(1e-10);
        for value in &mut embedding {
            *value /= norm;
        }

        if instrumented {
            DETERMINISTIC_PROVIDER_ENTRIES.fetch_sub(1, Ordering::SeqCst);
        }
        Ok(embedding)
    }

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

    fn model_name(&self) -> &str {
        "deterministic"
    }
}