mnemefusion-core 0.1.4

//! Configuration types for MnemeFusion
//!
//! This module defines configuration options for the memory engine.

use crate::query::FusionStrategy;

/// Configuration for the MnemeFusion memory engine
#[derive(Debug, Clone)]
pub struct Config {
    /// Dimension of embedding vectors
    pub embedding_dim: usize,

    /// Half-life for temporal decay in hours
    /// After this many hours, a memory's temporal relevance score is halved
    pub temporal_decay_hours: f32,

    /// Maximum number of hops for causal graph traversal
    pub causal_max_hops: usize,

    /// Enable automatic entity extraction
    ///
    /// **Language Note**: Entity extraction currently uses English-only stop words
    /// and capitalization rules. For non-English content, consider disabling this
    /// feature and using your own NER pipeline or relying on semantic search.
    ///
    /// See documentation for multilingual usage examples.
    pub entity_extraction_enabled: bool,

    /// Minimum confidence threshold for causal links (0.0 to 1.0)
    pub causal_min_confidence: f32,

    /// HNSW M parameter (connectivity)
    /// Higher values = better recall, more memory
    pub hnsw_m: usize,

    /// HNSW ef_construction parameter
    /// Higher values = better index quality, slower construction
    pub hnsw_ef_construction: usize,

    /// HNSW ef_search parameter
    /// Higher values = better recall, slower search
    pub hnsw_ef_search: usize,

    /// Metadata fields to index for efficient filtering
    /// These fields will have dedicated indexes for fast lookup
    /// Example: vec!["type".to_string(), "category".to_string(), "priority".to_string()]
    pub indexed_metadata: Vec<String>,

    /// Minimum semantic similarity threshold for fusion results (0.0 to 1.0)
    ///
    /// Memories with semantic_score below this threshold are excluded from 4D fusion results.
    /// This ensures that semantic relevance is mandatory - other dimensions (temporal, entity,
    /// causal) can only boost already-relevant memories, not surface irrelevant ones.
    ///
    /// Default: 0.15 (15% minimum semantic relevance)
    /// Recommended range: 0.10 to 0.20
    ///
    /// Lower values (e.g., 0.05): More permissive, may include weakly relevant memories
    /// Higher values (e.g., 0.30): Stricter, only strongly relevant memories
    ///
    /// Set to 0.0 to disable the filter (not recommended for production)
    pub fusion_semantic_threshold: f32,

    /// Pre-fusion semantic filtering threshold (0.0 to 1.0)
    ///
    /// Semantic search results below this threshold are filtered OUT before fusion.
    /// This is stricter than fusion_semantic_threshold and reduces noise in the semantic pathway.
    ///
    /// Default: 0.3 (30% minimum cosine similarity)
    /// Recommended range: 0.20 to 0.40
    ///
    /// This helps improve precision by removing low-quality semantic matches early.
    /// Set to 0.0 to disable pre-fusion filtering (uses only fusion_semantic_threshold).
    pub semantic_prefilter_threshold: f32,

    /// Fusion strategy (Weighted or ReciprocalRank)
    ///
    /// - Weighted: Uses intent-adaptive weights (original approach)
    /// - ReciprocalRank: Uses RRF formula (robust to score scale differences)
    ///
    /// Default: ReciprocalRank
    pub fusion_strategy: FusionStrategy,

    /// RRF k parameter (only used when fusion_strategy is ReciprocalRank)
    ///
    /// Default: 60 (from Cormack et al. 2009 RRF paper)
    pub rrf_k: f32,

    /// SLM configuration (optional)
    ///
    /// When enabled, uses Small Language Model for semantic intent classification
    /// instead of pattern matching. Improves classification accuracy from ~35% to 85%+.
    ///
    /// Default: None (disabled)
    ///
    /// Requires `slm` feature to be enabled at compile time:
    /// ```toml
    /// mnemefusion-core = { version = "0.1", features = ["slm"] }
    /// ```
    pub slm_config: Option<crate::slm::SlmConfig>,

    /// Enable SLM metadata extraction at ingestion time
    ///
    /// When enabled and slm_config is set, uses Small Language Model to extract rich
    /// metadata (entities, temporal markers, causal relationships, topics) during
    /// memory ingestion. This enables fast, accurate retrieval without query-time
    /// SLM inference.
    ///
    /// Key principle: "Pay the cost once" - ingestion can be slow (3-5s), but queries
    /// must be fast (<100ms).
    ///
    /// Default: true (enabled when slm_config is set)
    ///
    /// Requires `slm` feature to be enabled at compile time:
    /// ```toml
    /// mnemefusion-core = { version = "0.1", features = ["slm"] }
    /// ```
    pub slm_metadata_extraction_enabled: bool,

    /// Enable SLM query classification at query time
    ///
    /// When enabled and slm_config is set, uses Small Language Model to classify
    /// query intent at query time. This adds ~3 seconds latency to every query.
    ///
    /// **IMPORTANT**: This is typically NOT recommended for production use.
    /// With rich metadata extracted at ingestion time, RRF fusion
    /// can automatically balance retrieval pathways without classification.
    ///
    /// Default: false (disabled - rely on RRF fusion instead)
    ///
    /// Only enable for:
    /// - Experimentation and benchmarking
    /// - Comparing classification vs no-classification approaches
    /// - Cases where ingestion-time metadata is unavailable
    ///
    /// Requires `slm` feature to be enabled at compile time:
    /// ```toml
    /// mnemefusion-core = { version = "0.1", features = ["slm"] }
    /// ```
    pub slm_query_classification_enabled: bool,

    /// Number of LLM extraction passes per document during ingestion
    ///
    /// Multiple passes with different temperatures capture different facts,
    /// producing richer entity profiles. Pass 0 uses deterministic settings
    /// (temp=0.1), subsequent passes use moderate diversity (temp=0.3) with
    /// unique seeds.
    ///
    /// Default: 1 (single pass, backward compatible)
    /// Recommended for quality: 3
    ///
    /// Higher values increase ingestion time linearly but improve profile
    /// completeness. Facts are deduplicated across passes via add_fact().
    pub extraction_passes: usize,

    /// Entity types that get their own profiles during extraction.
    ///
    /// Only entities whose `entity_type` (from LLM extraction) matches one of
    /// these types (case-insensitive) will have profiles created. Others are
    /// skipped to prevent junk profiles from generic nouns, pronouns, and
    /// concepts (e.g., "dogs", "They", "basketball").
    ///
    /// Default: `["person", "organization", "location"]`
    ///
    /// Set to empty to allow all entity types (not recommended — causes
    /// profile explosion with diverse extraction prompts).
    pub profile_entity_types: Vec<String>,

    /// Adaptive-K threshold for dynamic result count selection (Top-p / nucleus).
    ///
    /// Instead of always returning exactly `limit` results, applies nucleus (Top-p)
    /// selection on fused scores: converts scores to probabilities via softmax, then
    /// returns results until cumulative probability >= this threshold.
    ///
    /// This prevents low-quality padding from diluting the context window. When the
    /// top few results have high scores and the rest are noise, Adaptive-K returns
    /// only the high-confidence results.
    ///
    /// Research basis: Calvin Ku's Adaptive-K fork of EmergenceMem Simple Fast.
    /// Top-p nucleus selection (p=0.7) with token budget. Average k varies per query.
    ///
    /// Default: 0.0 (disabled — always returns exactly `limit` results)
    /// Recommended: 0.7 (return results covering 70% of probability mass)
    /// Set to 0.0 to disable (always return `limit` results)
    /// Set to 1.0 to always return all results (equivalent to disabled)
    ///
    /// When enabled, results are bounded by [limit/3, limit] to prevent
    /// pathological cases (too few or too many results).
    pub adaptive_k_threshold: f32,

    /// Path to a local embedding model directory for automatic text vectorization.
    ///
    /// When set, `add()` and `query()` can be called without supplying an embedding
    /// vector — the engine computes it automatically using fastembed (ONNX Runtime).
    ///
    /// The path should be the root of a fastembed/HF-hub cache directory containing
    /// BGE-base-en-v1.5 model files. Pre-download with:
    /// `python -c "from fastembed import TextEmbedding; TextEmbedding(['hi'])"`
    ///
    /// Requires the `embedding-onnx` feature at compile time.
    ///
    /// Default: `None` (caller must supply embeddings explicitly)
    pub embedding_model: Option<String>,

    /// Path to the LLM model file (.gguf) for entity extraction.
    ///
    /// Convenience alias — equivalent to calling
    /// `engine.with_llm_entity_extraction_from_path(path, ModelTier::Balanced)` after `open()`.
    ///
    /// Requires the `entity-extraction` feature at compile time.
    ///
    /// Default: `None` (entity extraction disabled unless wired separately)
    pub llm_model: Option<String>,

    /// Minimum content size (bytes) to trigger deferred LLM extraction.
    ///
    /// When LLM extraction is enabled and `content.len() >= async_extraction_threshold`,
    /// `add()` stores the memory immediately and queues LLM extraction for later.
    /// Call `flush_extraction_queue()` to process all deferred extractions.
    ///
    /// This allows `add()` to return in ~5ms for large content (instead of ~9s per
    /// document), enabling real-time AI agent usage without blocking conversation flow.
    ///
    /// Default: `0` (always synchronous — backward compatible).
    /// Recommended: `500` (defer documents >= 500 bytes; short messages still sync).
    /// Set to `1` to defer all content when LLM is enabled.
    pub async_extraction_threshold: usize,

    /// Enable pipeline tracing (query).
    ///
    /// When enabled, each `query()` call records a detailed step-by-step
    /// trace accessible via `last_query_trace()`.
    ///
    /// Adds minor overhead from HashMap allocations (~5-10%). Disabled by default.
    ///
    /// Default: `false`
    pub enable_trace: bool,
}

impl Default for Config {
    fn default() -> Self {
        Self {
            embedding_dim: 384,          // Default for all-MiniLM-L6-v2
            temporal_decay_hours: 168.0, // 1 week
            causal_max_hops: 3,
            entity_extraction_enabled: true,
            causal_min_confidence: 0.5,
            hnsw_m: 16,
            hnsw_ef_construction: 128,
            hnsw_ef_search: 64,
            indexed_metadata: Vec::new(), // No indexed fields by default
            fusion_semantic_threshold: 0.15, // 15% minimum semantic relevance
            semantic_prefilter_threshold: 0.3, // 30% pre-fusion filter; aggregation queries use 0.5x (0.15) via adaptive logic in planner
            fusion_strategy: FusionStrategy::default(), // RRF by default
            rrf_k: 60.0,                       // From RRF paper
            slm_config: None,                  // SLM disabled by default
            slm_metadata_extraction_enabled: true, // Enabled by default when slm_config is set
            slm_query_classification_enabled: false, // Disabled by default - rely on RRF fusion
            extraction_passes: 1,              // Single pass by default (backward compatible)
            profile_entity_types: vec![
                "person".to_string(),
                "organization".to_string(),
                "location".to_string(),
            ],
            adaptive_k_threshold: 0.0, // Disabled by default (always return exactly limit)
            embedding_model: None,
            llm_model: None,
            async_extraction_threshold: 0, // Always sync by default (backward compatible)
            enable_trace: false,           // Pipeline tracing off by default
        }
    }
}

impl Config {
    /// Create a new configuration with default values
    pub fn new() -> Self {
        Self::default()
    }

    /// Set the embedding dimension
    pub fn with_embedding_dim(mut self, dim: usize) -> Self {
        self.embedding_dim = dim;
        self
    }

    /// Set the temporal decay half-life in hours
    pub fn with_temporal_decay_hours(mut self, hours: f32) -> Self {
        self.temporal_decay_hours = hours;
        self
    }

    /// Set the maximum causal traversal hops
    pub fn with_causal_max_hops(mut self, hops: usize) -> Self {
        self.causal_max_hops = hops;
        self
    }

    /// Enable or disable entity extraction
    ///
    /// **Language Note**: Entity extraction currently supports English only.
    /// If you're working with non-English content, set this to `false` and
    /// use multilingual embeddings for semantic search.
    ///
    /// # Examples
    ///
    /// ```
    /// use mnemefusion_core::Config;
    ///
    /// // For English content (default)
    /// let config = Config::new().with_entity_extraction(true);
    ///
    /// // For non-English content
    /// let config = Config::new().with_entity_extraction(false);
    /// ```
    pub fn with_entity_extraction(mut self, enabled: bool) -> Self {
        self.entity_extraction_enabled = enabled;
        self
    }

    /// Set HNSW parameters for vector index
    pub fn with_hnsw_params(mut self, m: usize, ef_construction: usize, ef_search: usize) -> Self {
        self.hnsw_m = m;
        self.hnsw_ef_construction = ef_construction;
        self.hnsw_ef_search = ef_search;
        self
    }

    /// Set metadata fields to index for efficient filtering
    pub fn with_indexed_metadata(mut self, fields: Vec<String>) -> Self {
        self.indexed_metadata = fields;
        self
    }

    /// Add a metadata field to the indexed set
    pub fn add_indexed_field(mut self, field: impl Into<String>) -> Self {
        self.indexed_metadata.push(field.into());
        self
    }

    /// Set the minimum semantic similarity threshold for fusion results
    ///
    /// Memories with semantic_score below this threshold are excluded from 4D fusion results.
    /// This ensures that semantic relevance is mandatory.
    ///
    /// # Arguments
    ///
    /// * `threshold` - Minimum semantic score (0.0 to 1.0). Default: 0.15
    ///
    /// # Examples
    ///
    /// ```
    /// use mnemefusion_core::Config;
    ///
    /// // Strict filter (only highly relevant memories)
    /// let config = Config::default().with_fusion_semantic_threshold(0.30);
    ///
    /// // Permissive filter (allow weakly relevant memories)
    /// let config = Config::default().with_fusion_semantic_threshold(0.05);
    ///
    /// // Disable filter (not recommended for production)
    /// let config = Config::default().with_fusion_semantic_threshold(0.0);
    /// ```
    pub fn with_fusion_semantic_threshold(mut self, threshold: f32) -> Self {
        self.fusion_semantic_threshold = threshold.clamp(0.0, 1.0);
        self
    }

    /// Set the fusion strategy
    ///
    /// # Arguments
    ///
    /// * `strategy` - Either Weighted or ReciprocalRank
    ///
    /// # Examples
    ///
    /// ```
    /// use mnemefusion_core::{Config, query::FusionStrategy};
    ///
    /// // Use RRF (default, recommended)
    /// let config = Config::default().with_fusion_strategy(FusionStrategy::ReciprocalRank);
    ///
    /// // Use weighted fusion (original approach)
    /// let config = Config::default().with_fusion_strategy(FusionStrategy::Weighted);
    /// ```
    pub fn with_fusion_strategy(mut self, strategy: FusionStrategy) -> Self {
        self.fusion_strategy = strategy;
        self
    }

    /// Set the RRF k parameter
    ///
    /// Only used when fusion_strategy is ReciprocalRank.
    ///
    /// # Arguments
    ///
    /// * `k` - RRF constant (typically 60). Default: 60
    ///
    /// # Examples
    ///
    /// ```
    /// use mnemefusion_core::Config;
    ///
    /// // Default k=60 (from RRF paper)
    /// let config = Config::default();
    ///
    /// // Custom k value
    /// let config = Config::default().with_rrf_k(100.0);
    /// ```
    pub fn with_rrf_k(mut self, k: f32) -> Self {
        self.rrf_k = k.max(1.0);
        self
    }

    /// Set the pre-fusion semantic filter threshold
    ///
    /// Semantic search results below this threshold are filtered out before fusion.
    /// This is stricter than fusion_semantic_threshold and helps improve precision.
    ///
    /// # Arguments
    ///
    /// * `threshold` - Minimum cosine similarity (0.0 to 1.0). Default: 0.3
    ///
    /// # Examples
    ///
    /// ```
    /// use mnemefusion_core::Config;
    ///
    /// // Strict pre-filter (only high-quality semantic matches)
    /// let config = Config::default().with_semantic_prefilter_threshold(0.4);
    ///
    /// // Permissive pre-filter
    /// let config = Config::default().with_semantic_prefilter_threshold(0.2);
    ///
    /// // Disable pre-filter (use only fusion threshold)
    /// let config = Config::default().with_semantic_prefilter_threshold(0.0);
    /// ```
    pub fn with_semantic_prefilter_threshold(mut self, threshold: f32) -> Self {
        self.semantic_prefilter_threshold = threshold.clamp(0.0, 1.0);
        self
    }

    /// Enable SLM-based intent classification
    ///
    /// When enabled, uses Small Language Model (Gemma 3 1B) for semantic understanding
    /// of query intent, improving classification accuracy from ~35% to 85%+.
    ///
    /// Falls back to pattern-based classification on any error, ensuring zero regression.
    ///
    /// # Arguments
    ///
    /// * `slm_config` - SLM configuration including model ID and cache directory
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use mnemefusion_core::{Config, SlmConfig};
    ///
    /// // Enable SLM with default configuration
    /// let config = Config::default()
    ///     .with_slm(SlmConfig::default());
    ///
    /// // Enable SLM with custom model
    /// let config = Config::default()
    ///     .with_slm(SlmConfig::new("google/gemma-3-1b")
    ///         .with_timeout_ms(100)
    ///         .with_min_confidence(0.6));
    /// ```
    ///
    /// # Feature Flag
    ///
    /// Requires `slm` feature to be enabled:
    /// ```toml
    /// [dependencies]
    /// mnemefusion-core = { version = "0.1", features = ["slm"] }
    /// ```
    pub fn with_slm(mut self, slm_config: crate::slm::SlmConfig) -> Self {
        self.slm_config = Some(slm_config);
        self
    }

    /// Disable SLM classification (use pattern-based only)
    ///
    /// This is the default behavior. Use this method to explicitly disable SLM
    /// after it has been enabled.
    pub fn without_slm(mut self) -> Self {
        self.slm_config = None;
        self
    }

    /// Enable or disable SLM metadata extraction at ingestion time
    ///
    /// When enabled and slm_config is set, uses Small Language Model to extract
    /// rich metadata during memory ingestion. This metadata enables fast, accurate
    /// retrieval without query-time SLM inference.
    ///
    /// # Arguments
    ///
    /// * `enabled` - Whether to enable SLM metadata extraction
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use mnemefusion_core::{Config, SlmConfig};
    ///
    /// // Enable SLM for both classification and metadata extraction
    /// let config = Config::default()
    ///     .with_slm(SlmConfig::default())
    ///     .with_slm_metadata_extraction(true);
    ///
    /// // Enable SLM for classification only, skip metadata extraction
    /// let config = Config::default()
    ///     .with_slm(SlmConfig::default())
    ///     .with_slm_metadata_extraction(false);
    /// ```
    pub fn with_slm_metadata_extraction(mut self, enabled: bool) -> Self {
        self.slm_metadata_extraction_enabled = enabled;
        self
    }

    /// Enable or disable SLM query classification at query time
    ///
    /// When enabled and slm_config is set, uses Small Language Model to classify
    /// query intent. This adds ~3 seconds latency per query.
    ///
    /// **NOT RECOMMENDED** for production. With rich metadata from ingestion-time
    /// SLM extraction, RRF fusion handles pathway balancing automatically.
    ///
    /// # Arguments
    ///
    /// * `enabled` - Whether to enable SLM query classification
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use mnemefusion_core::{Config, SlmConfig};
    ///
    /// // Default: SLM classification disabled at query time (recommended)
    /// let config = Config::default()
    ///     .with_slm(SlmConfig::default());
    ///
    /// // Enable for experimentation only
    /// let config = Config::default()
    ///     .with_slm(SlmConfig::default())
    ///     .with_slm_query_classification(true);
    /// ```
    pub fn with_slm_query_classification(mut self, enabled: bool) -> Self {
        self.slm_query_classification_enabled = enabled;
        self
    }

    /// Set the number of LLM extraction passes per document
    ///
    /// Multiple passes with different temperatures produce richer entity profiles
    /// by capturing different facts each time. The first pass is deterministic;
    /// subsequent passes use moderate randomness.
    ///
    /// # Arguments
    ///
    /// * `passes` - Number of extraction passes (1..=10). Clamped to this range.
    ///
    /// # Examples
    ///
    /// ```
    /// use mnemefusion_core::Config;
    ///
    /// // Single pass (default, fastest)
    /// let config = Config::default().with_extraction_passes(1);
    ///
    /// // Three passes (recommended for quality)
    /// let config = Config::default().with_extraction_passes(3);
    /// ```
    pub fn with_extraction_passes(mut self, passes: usize) -> Self {
        self.extraction_passes = passes.clamp(1, 10);
        self
    }

    /// Set which entity types get their own profiles during extraction.
    ///
    /// Only entities matching these types (case-insensitive) will have profiles
    /// created. An empty list allows all entity types.
    ///
    /// # Arguments
    ///
    /// * `types` - List of allowed entity types (e.g., `["person", "organization"]`)
    pub fn with_profile_entity_types(mut self, types: Vec<String>) -> Self {
        self.profile_entity_types = types;
        self
    }

    /// Set the Adaptive-K (Top-p) threshold for dynamic result count selection.
    ///
    /// When > 0.0, applies nucleus selection to return only high-confidence results
    /// instead of always returning exactly `limit` results.
    ///
    /// # Arguments
    ///
    /// * `threshold` - Top-p threshold (0.0 to 1.0). 0.0 disables, 0.7 recommended.
    pub fn with_adaptive_k(mut self, threshold: f32) -> Self {
        self.adaptive_k_threshold = threshold.clamp(0.0, 1.0);
        self
    }

    /// Set the path to a local embedding model directory.
    ///
    /// When set, the engine automatically computes embeddings for `add()` and
    /// `query()` calls that do not supply an explicit embedding vector.
    ///
    /// Requires the `embedding-onnx` feature at compile time.
    pub fn with_embedding_model(mut self, path: impl Into<String>) -> Self {
        self.embedding_model = Some(path.into());
        self
    }

    /// Set the path to the LLM model file (.gguf) for entity extraction.
    ///
    /// Requires the `entity-extraction` feature at compile time.
    pub fn with_llm_model(mut self, path: impl Into<String>) -> Self {
        self.llm_model = Some(path.into());
        self
    }

    /// Set the content size threshold for deferred LLM extraction.
    ///
    /// Content >= `threshold` bytes will be stored immediately and queued for
    /// deferred LLM extraction. Call `flush_extraction_queue()` to process.
    /// Set to `0` to always run LLM extraction synchronously (default).
    pub fn with_async_extraction_threshold(mut self, threshold: usize) -> Self {
        self.async_extraction_threshold = threshold;
        self
    }

    /// Enable or disable pipeline tracing.
    ///
    /// When enabled, `query()` records step-by-step traces with timing and data
    /// at every pipeline stage. Access via `last_query_trace()`.
    pub fn with_trace(mut self, enabled: bool) -> Self {
        self.enable_trace = enabled;
        self
    }

    /// Validate the configuration
    ///
    /// Returns detailed errors if the configuration is invalid.
    ///
    /// **Note**: Also prints warnings to stderr for suboptimal configurations
    /// (e.g., entity extraction enabled for potentially non-English content).
    pub fn validate(&self) -> Result<(), crate::Error> {
        // Print warning for entity extraction (English-only feature)
        if self.entity_extraction_enabled {
            tracing::warn!(
                "Entity extraction is enabled. This feature currently supports English only. \
                For non-English content, consider disabling with .with_entity_extraction(false)"
            );
        }

        if self.embedding_dim == 0 {
            return Err(crate::Error::Configuration(
                "embedding_dim must be greater than 0. Common values: 384 (MiniLM), 768 (BERT), 1536 (OpenAI)".to_string(),
            ));
        }

        // Warn if dimension is unusually large
        if self.embedding_dim > 4096 {
            return Err(crate::Error::Configuration(
                format!("embedding_dim of {} is unusually large. This will consume significant memory. Typical values are 384-1536.", self.embedding_dim)
            ));
        }

        if self.temporal_decay_hours <= 0.0 {
            return Err(crate::Error::Configuration(
                "temporal_decay_hours must be positive. Recommended: 168.0 (1 week)".to_string(),
            ));
        }

        if self.causal_max_hops == 0 {
            return Err(crate::Error::Configuration(
                "causal_max_hops must be greater than 0. Recommended: 2-5".to_string(),
            ));
        }

        // Warn if hops is very large
        if self.causal_max_hops > 10 {
            return Err(crate::Error::Configuration(format!(
                "causal_max_hops of {} may be too large and cause slow queries. Recommended: 2-5",
                self.causal_max_hops
            )));
        }

        if self.causal_min_confidence < 0.0 || self.causal_min_confidence > 1.0 {
            return Err(crate::Error::Configuration(format!(
                "causal_min_confidence must be between 0.0 and 1.0, got {}",
                self.causal_min_confidence
            )));
        }

        if self.hnsw_m == 0 {
            return Err(crate::Error::Configuration(
                "hnsw_m must be greater than 0. Recommended: 12-48 (default: 16)".to_string(),
            ));
        }

        if self.hnsw_m > 100 {
            return Err(crate::Error::Configuration(format!(
                "hnsw_m of {} is very large and will consume excessive memory. Recommended: 12-48",
                self.hnsw_m
            )));
        }

        if self.hnsw_ef_construction < 10 {
            return Err(crate::Error::Configuration(
                "hnsw_ef_construction should be at least 10 for reasonable index quality. Recommended: 100-500".to_string(),
            ));
        }

        if self.hnsw_ef_search == 0 {
            return Err(crate::Error::Configuration(
                "hnsw_ef_search must be greater than 0. Recommended: 64-200".to_string(),
            ));
        }

        if self.extraction_passes == 0 || self.extraction_passes > 10 {
            return Err(crate::Error::Configuration(format!(
                "extraction_passes must be between 1 and 10, got {}. Recommended: 1-3",
                self.extraction_passes
            )));
        }

        Ok(())
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_default_config() {
        let config = Config::default();
        assert_eq!(config.embedding_dim, 384);
        assert_eq!(config.temporal_decay_hours, 168.0);
        assert_eq!(config.causal_max_hops, 3);
        assert!(config.entity_extraction_enabled);
    }

    #[test]
    fn test_config_builder() {
        let config = Config::new()
            .with_embedding_dim(512)
            .with_temporal_decay_hours(336.0)
            .with_causal_max_hops(5);

        assert_eq!(config.embedding_dim, 512);
        assert_eq!(config.temporal_decay_hours, 336.0);
        assert_eq!(config.causal_max_hops, 5);
    }

    #[test]
    fn test_config_validation() {
        let config = Config::default();
        assert!(config.validate().is_ok());

        let mut bad_config = Config::default();
        bad_config.embedding_dim = 0;
        assert!(bad_config.validate().is_err());

        let mut bad_config = Config::default();
        bad_config.temporal_decay_hours = -1.0;
        assert!(bad_config.validate().is_err());

        let mut bad_config = Config::default();
        bad_config.causal_min_confidence = 1.5;
        assert!(bad_config.validate().is_err());
    }

    #[test]
    fn test_indexed_metadata_config() {
        // Default should have no indexed fields
        let config = Config::default();
        assert!(config.indexed_metadata.is_empty());

        // Test with_indexed_metadata
        let config =
            Config::new().with_indexed_metadata(vec!["type".to_string(), "category".to_string()]);
        assert_eq!(config.indexed_metadata.len(), 2);
        assert!(config.indexed_metadata.contains(&"type".to_string()));
        assert!(config.indexed_metadata.contains(&"category".to_string()));

        // Test add_indexed_field
        let config = Config::new()
            .add_indexed_field("type")
            .add_indexed_field("priority");
        assert_eq!(config.indexed_metadata.len(), 2);
        assert!(config.indexed_metadata.contains(&"type".to_string()));
        assert!(config.indexed_metadata.contains(&"priority".to_string()));
    }

    #[test]
    fn test_config_validation_dimension_too_large() {
        let mut config = Config::default();
        config.embedding_dim = 5000;
        let err = config.validate().unwrap_err();
        assert!(matches!(err, crate::Error::Configuration(_)));
        assert!(err.to_string().contains("unusually large"));
    }

    #[test]
    fn test_config_validation_causal_hops_too_large() {
        let mut config = Config::default();
        config.causal_max_hops = 20;
        let err = config.validate().unwrap_err();
        assert!(matches!(err, crate::Error::Configuration(_)));
        assert!(err.to_string().contains("too large"));
    }

    #[test]
    fn test_config_validation_hnsw_m_too_large() {
        let mut config = Config::default();
        config.hnsw_m = 150;
        let err = config.validate().unwrap_err();
        assert!(matches!(err, crate::Error::Configuration(_)));
        assert!(err.to_string().contains("very large"));
    }

    #[test]
    fn test_config_validation_ef_construction_too_small() {
        let mut config = Config::default();
        config.hnsw_ef_construction = 5;
        let err = config.validate().unwrap_err();
        assert!(matches!(err, crate::Error::Configuration(_)));
        assert!(err.to_string().contains("at least 10"));
    }

    #[test]
    fn test_config_validation_ef_search_zero() {
        let mut config = Config::default();
        config.hnsw_ef_search = 0;
        let err = config.validate().unwrap_err();
        assert!(matches!(err, crate::Error::Configuration(_)));
    }

    #[test]
    fn test_config_validation_provides_recommendations() {
        // Test that error messages include recommendations
        let mut config = Config::default();
        config.embedding_dim = 0;
        let err = config.validate().unwrap_err();
        assert!(err.to_string().contains("Common values"));

        let mut config = Config::default();
        config.temporal_decay_hours = 0.0;
        let err = config.validate().unwrap_err();
        assert!(err.to_string().contains("Recommended"));
    }

    #[test]
    fn test_entity_extraction_warning() {
        // Entity extraction enabled should validate successfully but print warning
        let config = Config::default();
        assert!(config.entity_extraction_enabled); // Default is true

        // Should not error, just warn to stderr
        assert!(config.validate().is_ok());
    }

    #[test]
    fn test_entity_extraction_disabled_no_warning() {
        // Disabling entity extraction should validate without warnings
        let config = Config::new().with_entity_extraction(false);
        assert!(!config.entity_extraction_enabled);

        // Should validate successfully
        assert!(config.validate().is_ok());
    }

    #[test]
    fn test_slm_query_classification_default_disabled() {
        // SLM query classification should be disabled by default
        let config = Config::default();
        assert!(!config.slm_query_classification_enabled);
    }

    #[test]
    fn test_slm_query_classification_builder() {
        // Test enabling SLM query classification
        let config = Config::default().with_slm_query_classification(true);
        assert!(config.slm_query_classification_enabled);

        // Test explicitly disabling
        let config = Config::default().with_slm_query_classification(false);
        assert!(!config.slm_query_classification_enabled);
    }
}