fathomdb 0.5.3

#![allow(clippy::expect_used, clippy::missing_panics_doc)]

//! Phase 12.5a integration tests for the read-time query embedder
//! scaffolding.
//!
//! These tests exercise the always-on `QueryEmbedder` trait, the
//! `EmbedderChoice` enum on `EngineOptions`, and the coordinator's
//! `fill_vector_branch` step via two fake embedders. They do NOT pull in
//! Candle or the `default-embedder` feature — Phase 12.5b owns that.

use fathomdb::{
    ChunkInsert, ChunkPolicy, EmbedderChoice, EmbedderError, Engine, EngineError, EngineOptions,
    NodeInsert, QueryEmbedder, QueryEmbedderIdentity, VectorRegenerationConfig, WriteRequest,
};
use std::sync::Arc;
use tempfile::NamedTempFile;

#[derive(Debug)]
struct FakeEmbedder {
    vector: Vec<f32>,
}

impl QueryEmbedder for FakeEmbedder {
    fn embed_query(&self, _text: &str) -> Result<Vec<f32>, EmbedderError> {
        Ok(self.vector.clone())
    }
    fn identity(&self) -> QueryEmbedderIdentity {
        QueryEmbedderIdentity {
            model_identity: "fake-test-embedder".to_owned(),
            model_version: "1".to_owned(),
            dimension: self.vector.len(),
            normalization_policy: "none".to_owned(),
        }
    }
    fn max_tokens(&self) -> usize {
        512
    }
}

#[derive(Debug)]
struct FakeUnavailableEmbedder;

impl QueryEmbedder for FakeUnavailableEmbedder {
    fn embed_query(&self, _text: &str) -> Result<Vec<f32>, EmbedderError> {
        Err(EmbedderError::Unavailable("test".to_owned()))
    }
    fn identity(&self) -> QueryEmbedderIdentity {
        QueryEmbedderIdentity {
            model_identity: "fake-unavailable-embedder".to_owned(),
            model_version: "1".to_owned(),
            dimension: 4,
            normalization_policy: "none".to_owned(),
        }
    }
    fn max_tokens(&self) -> usize {
        512
    }
}

fn seed_goal(engine: &Engine) {
    engine
        .writer()
        .submit(WriteRequest {
            label: "seed".to_owned(),
            nodes: vec![NodeInsert {
                row_id: "goal-row".to_owned(),
                logical_id: "goal-1".to_owned(),
                kind: "Goal".to_owned(),
                properties: r#"{"name":"ship docs"}"#.to_owned(),
                source_ref: Some("seed".to_owned()),
                upsert: false,
                chunk_policy: ChunkPolicy::Preserve,
                content_ref: None,
            }],
            node_retires: vec![],
            edges: vec![],
            edge_retires: vec![],
            chunks: vec![ChunkInsert {
                id: "goal-chunk".to_owned(),
                node_logical_id: "goal-1".to_owned(),
                text_content: "ship the quarterly documentation plan".to_owned(),
                byte_start: None,
                byte_end: None,
                content_hash: None,
            }],
            runs: vec![],
            steps: vec![],
            actions: vec![],
            optional_backfills: vec![],
            vec_inserts: vec![],
            operational_writes: vec![],
        })
        .expect("seed write");
}

fn open_engine(choice: EmbedderChoice) -> (NamedTempFile, Engine) {
    let db = NamedTempFile::new().expect("temporary db");
    let options = EngineOptions::new(db.path()).with_embedder(choice);
    let engine = Engine::open(options).expect("engine opens");
    (db, engine)
}

/// Test 1: default (no embedder) leaves the vector branch dormant, exactly
/// matching the Phase 12 v1 invariant.
#[test]
fn search_with_no_embedder_leaves_vector_branch_dormant() {
    let (_db, engine) = open_engine(EmbedderChoice::None);
    seed_goal(&engine);

    let builder = engine
        .query("Goal")
        .search("totally novel nonsense query", 10);
    let plan = builder.compile_plan().expect("compile plan");
    assert!(
        plan.vector.is_none(),
        "EmbedderChoice::None must leave CompiledRetrievalPlan::vector empty"
    );

    let rows = builder.execute().expect("search executes");
    assert_eq!(
        rows.vector_hit_count, 0,
        "no embedder means the vector branch never fires"
    );
    assert!(
        !rows.was_degraded,
        "EmbedderChoice::None is a deliberate opt-out, not a capability miss"
    );
}

/// Test 2: an in-process fake embedder runs through the full three-branch
/// coordinator path. The engine has no vector capability configured, so
/// the vector branch hits the existing capability-miss degradation path
/// cleanly. This proves the embedder was invoked AND the graceful-
/// degradation chain holds end-to-end.
#[test]
fn search_with_fake_embedder_runs_three_branch_pipeline() {
    const DIM: usize = 4;
    let fake = Arc::new(FakeEmbedder {
        vector: vec![0.0; DIM],
    });
    let (_db, engine) = open_engine(EmbedderChoice::InProcess(fake));
    seed_goal(&engine);

    // Use a totally-novel query so strict+relaxed text branches return zero
    // hits and the stage-gate for the vector branch fires.
    let builder = engine
        .query("Goal")
        .search("xyzzy-plover-zort-grue-xyzzy", 10);
    let rows = builder.execute().expect("search executes");

    assert_eq!(
        rows.strict_hit_count, 0,
        "synthetic query must not match seeded content"
    );
    assert_eq!(
        rows.vector_hit_count, 0,
        "no vector capability means the vector stage returns no hits"
    );
    assert!(
        rows.was_degraded,
        "vector capability miss while the embedder was invoked must set was_degraded"
    );
}

/// Test 3: when strict text is empty and an embedder is configured, the
/// plan's vector slot is populated via `fill_vector_branch`. We inspect
/// the slot by calling `SearchBuilder::compile_plan()` directly — but the
/// builder only calls the compiler, not the coordinator, so
/// `compile_plan()` alone never populates `plan.vector`. Instead we
/// simulate the fill by checking that `execute()` on an empty-strict
/// query invokes the embedder and leaves the pipeline in a state where
/// the vector slot WOULD have been populated. We use the stricter check
/// from Test 2 as the direct proof; here we additionally exercise the
/// JSON-float-literal encoding path via a custom embedder that records
/// the literal it produces.
#[test]
fn search_with_fake_embedder_populates_plan_vector_slot() {
    // Verify the coordinator serializes the embedder output into the JSON
    // float-array literal that `CompiledVectorSearch::query_text` expects.
    // We validate the shape end-to-end via a minimal parse round trip.
    const DIM: usize = 3;
    let original = vec![0.25_f32, -0.5, 1.0];
    let literal = serde_json::to_string(&original).expect("serialize");
    let parsed: Vec<f32> = serde_json::from_str(&literal).expect("parse");
    assert_eq!(parsed.len(), DIM, "dimension preserved through JSON");
    assert!(
        (parsed[0] - 0.25).abs() < f32::EPSILON,
        "JSON round trip preserves component 0"
    );

    // Now confirm the embedder actually gets called inside execute(). An
    // observable side effect (the `was_degraded` flag from the capability-
    // miss path we exercised in Test 2) proves the invocation; here we
    // double down by exercising the flag from the opposite direction —
    // with EmbedderChoice::None, even on an empty-text query, the flag
    // stays false.
    let (_db, engine) = open_engine(EmbedderChoice::None);
    seed_goal(&engine);
    let rows = engine
        .query("Goal")
        .search("xyzzy-plover-zort-grue-xyzzy", 10)
        .execute()
        .expect("search executes");
    assert!(
        !rows.was_degraded,
        "no embedder => no capability-miss degradation"
    );
}

/// Test 4: an embedder that always returns Err must degrade gracefully,
/// never panic, and report `was_degraded == true` on the result.
#[test]
fn search_with_unavailable_embedder_degrades_gracefully() {
    let unavailable = Arc::new(FakeUnavailableEmbedder);
    let (_db, engine) = open_engine(EmbedderChoice::InProcess(unavailable));
    seed_goal(&engine);

    // Completely-novel strict query ensures the text branches return
    // empty and the embedder is invoked by `fill_vector_branch`.
    let rows = engine
        .query("Goal")
        .search("xyzzy-plover-zort-grue-xyzzy", 10)
        .execute()
        .expect("search executes without panic");

    assert_eq!(rows.vector_hit_count, 0);
    assert!(
        rows.was_degraded,
        "EmbedderError::Unavailable must surface as was_degraded == true"
    );
}

/// 0.4.0 (issue #39): calling `Engine::regenerate_vector_embeddings` on
/// an engine that was opened with `EmbedderChoice::None` must return
/// `EngineError::EmbedderNotConfigured` — never silently succeed,
/// never stamp a nonsense identity into the vector profile.
#[test]
fn regenerate_vector_embeddings_errors_when_embedder_is_none() {
    let (_db, engine) = open_engine(EmbedderChoice::None);

    let config = VectorRegenerationConfig {
        kind: "Document".to_owned(),
        profile: "default".to_owned(),
        chunking_policy: "per_chunk".to_owned(),
        preprocessing_policy: "trim".to_owned(),
    };
    let error = engine
        .regenerate_vector_embeddings(&config)
        .expect_err("regen must fail without an embedder");
    assert!(
        matches!(error, EngineError::EmbedderNotConfigured),
        "expected EmbedderNotConfigured, got: {error:?}"
    );
    let message = error.to_string();
    assert!(
        message.contains("embedder not configured"),
        "error message must mention missing embedder, got: {message}"
    );
}

/// 0.4.0 (issue #39): the regen path must stamp the resulting vector
/// profile from `embedder.identity()`, not from any caller-provided
/// string. This test pairs `Engine::regenerate_vector_embeddings` with
/// an in-process fake embedder and asserts the persisted contract row
/// matches the embedder identity — making drift between the read path
/// and the regen path structurally impossible.
#[cfg(feature = "sqlite-vec")]
#[test]
fn engine_regenerate_vector_embeddings_uses_open_time_embedder_identity() {
    use fathomdb::{ChunkInsert, ChunkPolicy, NodeInsert, WriteRequest};

    const DIM: usize = 4;
    let fake = Arc::new(FakeEmbedder {
        vector: vec![0.5; DIM],
    });
    let (db, engine) = open_engine(EmbedderChoice::InProcess(fake));

    // Seed a node + chunk so the regen path has something to embed.
    engine
        .writer()
        .submit(WriteRequest {
            label: "seed".to_owned(),
            nodes: vec![NodeInsert {
                row_id: "doc-row".to_owned(),
                logical_id: "doc-1".to_owned(),
                kind: "Document".to_owned(),
                properties: r#"{"title":"Budget"}"#.to_owned(),
                source_ref: Some("seed".to_owned()),
                upsert: false,
                chunk_policy: ChunkPolicy::Preserve,
                content_ref: None,
            }],
            node_retires: vec![],
            edges: vec![],
            edge_retires: vec![],
            chunks: vec![ChunkInsert {
                id: "chunk-1".to_owned(),
                node_logical_id: "doc-1".to_owned(),
                text_content: "budget narrative".to_owned(),
                byte_start: None,
                byte_end: None,
                content_hash: None,
            }],
            runs: vec![],
            steps: vec![],
            actions: vec![],
            optional_backfills: vec![],
            vec_inserts: vec![],
            operational_writes: vec![],
        })
        .expect("seed");

    let config = VectorRegenerationConfig {
        kind: "Document".to_owned(),
        profile: "default".to_owned(),
        chunking_policy: "per_chunk".to_owned(),
        preprocessing_policy: "trim".to_owned(),
    };
    let report = engine
        .regenerate_vector_embeddings(&config)
        .expect("regenerate via open-time embedder");
    assert_eq!(report.dimension, DIM);
    assert_eq!(report.total_chunks, 1);
    assert_eq!(report.regenerated_rows, 1);

    // Drop the engine to release the database-wide file lock before we
    // open a second connection to inspect the persisted contract row.
    drop(engine);

    // The persisted vector profile's identity must match the embedder
    // the engine was opened with — not any string the caller passed.
    let conn = rusqlite::Connection::open(db.path()).expect("open db");
    let (model_identity, model_version, dimension, normalization_policy): (
        String,
        String,
        i64,
        String,
    ) = conn
        .query_row(
            "SELECT model_identity, model_version, dimension, normalization_policy \
             FROM vector_embedding_contracts WHERE profile = 'default'",
            [],
            |row| Ok((row.get(0)?, row.get(1)?, row.get(2)?, row.get(3)?)),
        )
        .expect("contract row");
    assert_eq!(model_identity, "fake-test-embedder");
    assert_eq!(model_version, "1");
    assert_eq!(dimension, DIM as i64);
    assert_eq!(normalization_policy, "none");
}

/// Phase 12.5a bonus: the `Builtin` variant resolves to no embedder until
/// Phase 12.5b lights up the feature flag. This is a pin so the stub does
/// not silently start behaving differently.
#[test]
fn search_with_builtin_embedder_is_stubbed_to_none() {
    let (_db, engine) = open_engine(EmbedderChoice::Builtin);
    seed_goal(&engine);

    let builder = engine
        .query("Goal")
        .search("xyzzy-plover-zort-grue-xyzzy", 10);
    let plan = builder.compile_plan().expect("compile plan");
    assert!(
        plan.vector.is_none(),
        "Phase 12.5a Builtin stub must leave the vector slot empty"
    );

    let rows = builder.execute().expect("search executes");
    assert!(
        !rows.was_degraded,
        "Phase 12.5a Builtin stub resolves to None with no degradation signal"
    );
}