quiverdb-core 0.28.0

// SPDX-License-Identifier: AGPL-3.0-only
//! Collection descriptors: the schema fixed when a collection is created.

use serde::{Deserialize, Serialize};

/// The element type of stored vectors. Phase 1 ships `f32`; lower-precision and
/// quantized dtypes arrive with the memory-frugality work in Phase 2.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[non_exhaustive]
pub enum Dtype {
    /// 32-bit IEEE-754 float.
    F32,
}

impl Dtype {
    /// Size in bytes of one vector element.
    #[must_use]
    pub const fn element_size(self) -> usize {
        match self {
            Dtype::F32 => 4,
        }
    }
}

/// The distance / similarity function a collection is searched with.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
pub enum DistanceMetric {
    /// Inner product — higher is more similar.
    Dot,
    /// Cosine similarity — higher is more similar.
    Cosine,
    /// Squared Euclidean distance — lower is more similar.
    L2,
}

/// The index structure a collection is served by (ADR-0007). The default is the
/// in-memory HNSW graph; the others are the Phase 2 memory-frugal options.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
#[non_exhaustive]
pub enum IndexKind {
    /// In-memory HNSW graph: lowest latency, fits in RAM. The default.
    #[default]
    Hnsw,
    /// In-memory Vamana (DiskANN) graph.
    Vamana,
    /// Disk-resident Vamana: PQ codes in RAM, graph + full vectors on SSD.
    DiskVamana,
    /// Inverted-file index with coarse Voronoi partitioning.
    Ivf,
    /// ColBERTv2/PLAID compressed token-pool index for late-interaction
    /// (`multivector`) collections: centroid + residual-PQ codes in RAM with
    /// centroid-pruned candidate generation (ADR-0034). Valid only for a
    /// `multivector` collection.
    Colbert,
}

/// Which index a collection uses and how its vectors are compressed (ADR-0007,
/// ADR-0008). Defaults to in-memory HNSW with no quantization (exact search).
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
pub struct IndexSpec {
    /// The index structure.
    pub kind: IndexKind,
    /// Product-quantization subspaces for quantized kinds (the disk graph,
    /// IVF+PQ). `None` selects a kind-appropriate default or no quantization.
    pub pq_subspaces: Option<u32>,
}

/// The type of a filterable payload field, which fixes how its values are keyed
/// in the secondary index (`.sec`) — and therefore which predicates it answers.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
#[non_exhaustive]
pub enum FieldType {
    /// An exact-match string field (equality and lexical range).
    Keyword,
    /// A numeric field (equality and numeric range), keyed order-preserving.
    Numeric,
}

/// A payload field declared filterable at collection creation: its dot-path and
/// type. Declared fields are extracted into the per-segment secondary index at
/// flush time (ADR-0022), enabling pre-filtered (hybrid) search.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct FilterableField {
    /// Dot-path into the JSON payload (e.g. `"user.age"`).
    pub path: String,
    /// The field's value type.
    pub field_type: FieldType,
}

impl FilterableField {
    /// A keyword (exact-match string) field at `path`.
    #[must_use]
    pub fn keyword(path: impl Into<String>) -> Self {
        Self {
            path: path.into(),
            field_type: FieldType::Keyword,
        }
    }

    /// A numeric field at `path`.
    #[must_use]
    pub fn numeric(path: impl Into<String>) -> Self {
        Self {
            path: path.into(),
            field_type: FieldType::Numeric,
        }
    }
}

/// How a collection's vectors are encrypted (ADR-0031, ADR-0032). Encryption is
/// always **client-side** — the server never holds the key. Defaults to
/// [`VectorEncryption::None`]. The variants sit on Quiver's encrypted-search
/// spectrum, from fastest to most confidential:
///
/// - [`None`](VectorEncryption::None): plaintext vectors; the server ranks and
///   sees everything (the default).
/// - [`Dcpe`](VectorEncryption::Dcpe): experimental property-preserving
///   encryption; the server ranks ciphertexts but the approximate
///   distance-comparison relation leaks **by design**. `L2` only; not
///   semantically secure (ADR-0031).
/// - [`ClientSide`](VectorEncryption::ClientSide): semantically secure (IND-CPA)
///   opaque AEAD ciphertext; the server stores blobs it cannot read and does
///   **no** distance math, so the client fetches and ranks locally (ADR-0032).
///
/// The discriminants are chosen so a descriptor written when this flag was a
/// `bool encrypted_vectors` decodes unchanged: `false` (byte 0) is
/// [`None`](VectorEncryption::None) and `true` (byte 1) is
/// [`Dcpe`](VectorEncryption::Dcpe) — no data migration.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
pub enum VectorEncryption {
    /// Plaintext vectors; the server ranks (the default).
    #[default]
    None,
    /// Experimental DCPE ciphertext (ADR-0031): the server ranks, the approximate
    /// distance-comparison relation leaks by design, `L2` only, not semantically
    /// secure.
    Dcpe,
    /// Semantically secure opaque AEAD ciphertext (ADR-0032): the server stores
    /// blobs it cannot read and does no distance math; the client ranks locally.
    ClientSide,
}

/// The immutable schema of a collection, fixed at creation.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct Descriptor {
    /// Vector dimensionality.
    pub dim: u32,
    /// Vector element type.
    pub dtype: Dtype,
    /// Distance metric used for search.
    pub metric: DistanceMetric,
    /// Index & quantization configuration. Defaults to HNSW/exact and is absent
    /// in descriptors written before Phase 2 (filled by the default on read).
    #[serde(default)]
    pub index: IndexSpec,
    /// Payload fields indexed for filtering. Empty by default and absent in
    /// descriptors written before secondary indexes existed (defaulted on read).
    #[serde(default)]
    pub filterable: Vec<FilterableField>,
    /// Whether this is a multi-vector (late-interaction / ColBERT) collection:
    /// each document is stored as a group of token-vector rows and searched by
    /// MaxSim (ADR-0028). `false` for an ordinary single-vector collection, and
    /// absent in descriptors written before late interaction existed (defaulted to
    /// `false` on read).
    #[serde(default)]
    pub multivector: bool,
    /// How this collection's vectors are encrypted (ADR-0031, ADR-0032).
    /// [`VectorEncryption::None`] by default; [`Dcpe`](VectorEncryption::Dcpe)
    /// requires the `L2` metric. Absent in descriptors written before the flag
    /// existed (defaulted to `None` on read); a descriptor written while the flag
    /// was a `bool encrypted_vectors` decodes unchanged (`false`→`None`,
    /// `true`→`Dcpe`).
    #[serde(default)]
    pub vector_encryption: VectorEncryption,
}

impl Descriptor {
    /// A descriptor with the default index (in-memory HNSW, exact) and no
    /// filterable fields.
    #[must_use]
    pub fn new(dim: u32, dtype: Dtype, metric: DistanceMetric) -> Self {
        Self {
            dim,
            dtype,
            metric,
            index: IndexSpec::default(),
            filterable: Vec::new(),
            multivector: false,
            vector_encryption: VectorEncryption::None,
        }
    }

    /// Set the index specification (builder style).
    #[must_use]
    pub fn with_index(mut self, index: IndexSpec) -> Self {
        self.index = index;
        self
    }

    /// Set the filterable payload fields (builder style).
    #[must_use]
    pub fn with_filterable(mut self, filterable: Vec<FilterableField>) -> Self {
        self.filterable = filterable;
        self
    }

    /// Mark this collection as multi-vector (late-interaction / ColBERT), so each
    /// document is stored as a group of token-vector rows scored by MaxSim
    /// (builder style). The dimensionality is the per-token dimension.
    #[must_use]
    pub fn with_multivector(mut self, multivector: bool) -> Self {
        self.multivector = multivector;
        self
    }

    /// Set how this collection's vectors are encrypted (builder style). A
    /// [`Dcpe`](VectorEncryption::Dcpe) collection must use the `L2` metric; a
    /// [`ClientSide`](VectorEncryption::ClientSide) collection is searched by the
    /// client, not the server (ADR-0031, ADR-0032).
    #[must_use]
    pub fn with_vector_encryption(mut self, vector_encryption: VectorEncryption) -> Self {
        self.vector_encryption = vector_encryption;
        self
    }

    /// Decode a descriptor from its postcard bytes, tolerating every earlier
    /// layout.
    ///
    /// postcard is non-self-describing, so a missing *trailing* field cannot be
    /// defaulted by `#[serde(default)]` alone (the reader hits end-of-input and
    /// errors). We therefore try the layouts newest-to-oldest — current
    /// (with `vector_encryption`) → the six-field `multivector` layout → the
    /// five-field `filterable` layout → the four-field `index`-only layout → the
    /// original three-field layout — defaulting the missing trailing fields. The
    /// order matters: postcard ignores trailing bytes, so an older decoder would
    /// silently mis-read a newer buffer if tried first.
    ///
    /// # Errors
    /// Returns the postcard error if the bytes match no known layout.
    pub fn decode(bytes: &[u8]) -> std::result::Result<Self, postcard::Error> {
        postcard::from_bytes::<Self>(bytes)
            .or_else(|_| postcard::from_bytes::<DescriptorV4>(bytes).map(Self::from))
            .or_else(|_| postcard::from_bytes::<DescriptorV3>(bytes).map(Self::from))
            .or_else(|_| postcard::from_bytes::<DescriptorV2>(bytes).map(Self::from))
            .or_else(|_| postcard::from_bytes::<LegacyDescriptor>(bytes).map(Self::from))
    }

    /// Byte length of one stored vector (`dim × element_size`).
    #[must_use]
    pub fn stride(&self) -> usize {
        self.dim as usize * self.dtype.element_size()
    }
}

// The six-field layout (through `multivector`, no `vector_encryption`), kept only
// to migrate descriptors written before client-side encryption existed, via
// [`Descriptor::decode`].
// It must be tried before the five-field layout, which would otherwise silently
// drop `multivector` (postcard ignores trailing bytes).
#[derive(Deserialize)]
struct DescriptorV4 {
    dim: u32,
    dtype: Dtype,
    metric: DistanceMetric,
    index: IndexSpec,
    filterable: Vec<FilterableField>,
    multivector: bool,
}

impl From<DescriptorV4> for Descriptor {
    fn from(v: DescriptorV4) -> Self {
        Self {
            dim: v.dim,
            dtype: v.dtype,
            metric: v.metric,
            index: v.index,
            filterable: v.filterable,
            multivector: v.multivector,
            vector_encryption: VectorEncryption::None,
        }
    }
}

// The five-field layout (an `index` and `filterable` but no `multivector`), kept
// only to migrate descriptors written before late interaction existed, via
// [`Descriptor::decode`]. It must be tried before the four-field layout, which
// would otherwise silently drop `filterable` (postcard ignores trailing bytes).
#[derive(Deserialize)]
struct DescriptorV3 {
    dim: u32,
    dtype: Dtype,
    metric: DistanceMetric,
    index: IndexSpec,
    filterable: Vec<FilterableField>,
}

impl From<DescriptorV3> for Descriptor {
    fn from(v: DescriptorV3) -> Self {
        Self {
            dim: v.dim,
            dtype: v.dtype,
            metric: v.metric,
            index: v.index,
            filterable: v.filterable,
            multivector: false,
            vector_encryption: VectorEncryption::None,
        }
    }
}

// The four-field layout (an `index` but no `filterable`), kept only to migrate
// descriptors written before secondary indexes existed, via [`Descriptor::decode`].
#[derive(Deserialize)]
struct DescriptorV2 {
    dim: u32,
    dtype: Dtype,
    metric: DistanceMetric,
    index: IndexSpec,
}

impl From<DescriptorV2> for Descriptor {
    fn from(v: DescriptorV2) -> Self {
        Self {
            dim: v.dim,
            dtype: v.dtype,
            metric: v.metric,
            index: v.index,
            filterable: Vec::new(),
            multivector: false,
            vector_encryption: VectorEncryption::None,
        }
    }
}

// The original three-field layout (no `index`, no `filterable`), kept only to
// migrate the oldest databases on read via [`Descriptor::decode`].
#[derive(Deserialize)]
struct LegacyDescriptor {
    dim: u32,
    dtype: Dtype,
    metric: DistanceMetric,
}

impl From<LegacyDescriptor> for Descriptor {
    fn from(v: LegacyDescriptor) -> Self {
        Self {
            dim: v.dim,
            dtype: v.dtype,
            metric: v.metric,
            index: IndexSpec::default(),
            filterable: Vec::new(),
            multivector: false,
            vector_encryption: VectorEncryption::None,
        }
    }
}

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

    #[test]
    fn stride_matches_dim_and_dtype() {
        let d = Descriptor::new(128, Dtype::F32, DistanceMetric::L2);
        assert_eq!(d.stride(), 512);
        assert_eq!(Dtype::F32.element_size(), 4);
        // The default index is in-memory HNSW with no quantization.
        assert_eq!(d.index, IndexSpec::default());
        assert_eq!(d.index.kind, IndexKind::Hnsw);
    }

    #[test]
    fn descriptor_roundtrips_through_postcard() {
        let d = Descriptor::new(8, Dtype::F32, DistanceMetric::Cosine).with_index(IndexSpec {
            kind: IndexKind::DiskVamana,
            pq_subspaces: Some(16),
        });
        let bytes = postcard::to_allocvec(&d).unwrap();
        let back: Descriptor = postcard::from_bytes(&bytes).unwrap();
        assert_eq!(d, back);
    }

    // A descriptor serialized before the `index` field existed (only dim, dtype,
    // metric) must still deserialize, defaulting the index to HNSW.
    #[test]
    fn pre_phase2_descriptor_deserializes_with_default_index() {
        #[derive(serde::Serialize)]
        struct OldDescriptor {
            dim: u32,
            dtype: Dtype,
            metric: DistanceMetric,
        }
        let old = OldDescriptor {
            dim: 16,
            dtype: Dtype::F32,
            metric: DistanceMetric::L2,
        };
        let bytes = postcard::to_allocvec(&old).unwrap();
        // The raw new-layout decode fails on the shorter legacy bytes...
        assert!(postcard::from_bytes::<Descriptor>(&bytes).is_err());
        // ...but `decode` falls back to the legacy layout and defaults the index.
        let back = Descriptor::decode(&bytes).unwrap();
        assert_eq!(back.dim, 16);
        assert_eq!(back.metric, DistanceMetric::L2);
        assert_eq!(back.index, IndexSpec::default());
    }

    #[test]
    fn decode_reads_current_layout() {
        let d = Descriptor::new(8, Dtype::F32, DistanceMetric::Dot).with_index(IndexSpec {
            kind: IndexKind::Ivf,
            pq_subspaces: Some(8),
        });
        let bytes = postcard::to_allocvec(&d).unwrap();
        assert_eq!(Descriptor::decode(&bytes).unwrap(), d);
    }

    // A descriptor serialized before `filterable` existed (four fields, with an
    // `index`) must still deserialize — and the four-field fallback must run
    // before the three-field one, so the `index` is preserved, not defaulted.
    #[test]
    fn pre_filterable_descriptor_decodes_and_keeps_its_index() {
        #[derive(serde::Serialize)]
        struct DescriptorV2 {
            dim: u32,
            dtype: Dtype,
            metric: DistanceMetric,
            index: IndexSpec,
        }
        let old = DescriptorV2 {
            dim: 8,
            dtype: Dtype::F32,
            metric: DistanceMetric::L2,
            index: IndexSpec {
                kind: IndexKind::DiskVamana,
                pq_subspaces: Some(16),
            },
        };
        let bytes = postcard::to_allocvec(&old).unwrap();
        // The current five-field decode fails on the shorter buffer...
        assert!(postcard::from_bytes::<Descriptor>(&bytes).is_err());
        // ...but `decode` falls back to the four-field layout, keeping the index
        // (not the three-field legacy layout, which would lose it).
        let back = Descriptor::decode(&bytes).unwrap();
        assert_eq!(back.dim, 8);
        assert_eq!(back.index.kind, IndexKind::DiskVamana);
        assert_eq!(back.index.pq_subspaces, Some(16));
        assert!(back.filterable.is_empty());
    }

    #[test]
    fn descriptor_with_filterable_roundtrips() {
        let d = Descriptor::new(4, Dtype::F32, DistanceMetric::L2).with_filterable(vec![
            FilterableField::keyword("city"),
            FilterableField::numeric("age"),
        ]);
        let bytes = postcard::to_allocvec(&d).unwrap();
        assert_eq!(Descriptor::decode(&bytes).unwrap(), d);
    }

    #[test]
    fn descriptor_with_multivector_roundtrips() {
        let d = Descriptor::new(128, Dtype::F32, DistanceMetric::Cosine).with_multivector(true);
        let bytes = postcard::to_allocvec(&d).unwrap();
        let back = Descriptor::decode(&bytes).unwrap();
        assert_eq!(back, d);
        assert!(back.multivector);
    }

    // A descriptor serialized before `multivector` existed (five fields, with a
    // `filterable`) must still decode — via the five-field fallback, which keeps
    // `filterable` and defaults `multivector` to false. The four-field fallback
    // would wrongly drop `filterable`, so the five-field one must be tried first.
    #[test]
    fn pre_multivector_descriptor_decodes_and_keeps_filterable() {
        #[derive(serde::Serialize)]
        struct DescriptorV3 {
            dim: u32,
            dtype: Dtype,
            metric: DistanceMetric,
            index: IndexSpec,
            filterable: Vec<FilterableField>,
        }
        let old = DescriptorV3 {
            dim: 8,
            dtype: Dtype::F32,
            metric: DistanceMetric::Cosine,
            index: IndexSpec {
                kind: IndexKind::Ivf,
                pq_subspaces: Some(8),
            },
            filterable: vec![FilterableField::keyword("city")],
        };
        let bytes = postcard::to_allocvec(&old).unwrap();
        // The current six-field decode fails on the shorter buffer...
        assert!(postcard::from_bytes::<Descriptor>(&bytes).is_err());
        // ...but `decode` falls back to the five-field layout: filterable kept,
        // multivector defaulted to false, index preserved.
        let back = Descriptor::decode(&bytes).unwrap();
        assert_eq!(back.filterable, vec![FilterableField::keyword("city")]);
        assert!(!back.multivector);
        assert_eq!(back.index.kind, IndexKind::Ivf);
    }

    #[test]
    fn descriptor_with_vector_encryption_roundtrips() {
        let d = Descriptor::new(64, Dtype::F32, DistanceMetric::L2)
            .with_vector_encryption(VectorEncryption::ClientSide);
        let bytes = postcard::to_allocvec(&d).unwrap();
        let back = Descriptor::decode(&bytes).unwrap();
        assert_eq!(back, d);
        assert_eq!(back.vector_encryption, VectorEncryption::ClientSide);
    }

    // The vector-encryption flag used to be a `bool encrypted_vectors`. The enum's
    // discriminants are chosen so those descriptors decode unchanged — a trailing
    // `true` byte is `Dcpe`, `false` is `None` — so existing DCPE collections need
    // no data migration.
    #[test]
    fn legacy_encrypted_vectors_bool_decodes_as_the_enum() {
        #[derive(serde::Serialize)]
        struct OldDescriptor {
            dim: u32,
            dtype: Dtype,
            metric: DistanceMetric,
            index: IndexSpec,
            filterable: Vec<FilterableField>,
            multivector: bool,
            encrypted_vectors: bool,
        }
        let make = |enc: bool| OldDescriptor {
            dim: 8,
            dtype: Dtype::F32,
            metric: DistanceMetric::L2,
            index: IndexSpec::default(),
            filterable: Vec::new(),
            multivector: false,
            encrypted_vectors: enc,
        };
        let dcpe = postcard::to_allocvec(&make(true)).unwrap();
        assert_eq!(
            Descriptor::decode(&dcpe).unwrap().vector_encryption,
            VectorEncryption::Dcpe
        );
        let none = postcard::to_allocvec(&make(false)).unwrap();
        assert_eq!(
            Descriptor::decode(&none).unwrap().vector_encryption,
            VectorEncryption::None
        );
    }

    // A descriptor serialized before `vector_encryption` existed (six fields,
    // through `multivector`) must still decode — via the six-field fallback, which
    // keeps `multivector` and defaults `vector_encryption` to None. The five-field
    // fallback would wrongly drop `multivector`, so the six-field one is tried first.
    #[test]
    fn pre_vector_encryption_descriptor_decodes_and_keeps_multivector() {
        #[derive(serde::Serialize)]
        struct DescriptorV4 {
            dim: u32,
            dtype: Dtype,
            metric: DistanceMetric,
            index: IndexSpec,
            filterable: Vec<FilterableField>,
            multivector: bool,
        }
        let old = DescriptorV4 {
            dim: 8,
            dtype: Dtype::F32,
            metric: DistanceMetric::Cosine,
            index: IndexSpec::default(),
            filterable: vec![FilterableField::numeric("score")],
            multivector: true,
        };
        let bytes = postcard::to_allocvec(&old).unwrap();
        // The current seven-field decode fails on the shorter buffer...
        assert!(postcard::from_bytes::<Descriptor>(&bytes).is_err());
        // ...but `decode` falls back to the six-field layout: multivector and
        // filterable kept, vector_encryption defaulted to None.
        let back = Descriptor::decode(&bytes).unwrap();
        assert!(back.multivector);
        assert_eq!(back.filterable, vec![FilterableField::numeric("score")]);
        assert_eq!(back.vector_encryption, VectorEncryption::None);
    }
}