fix: address PR review comments (round 2)

Human reviewer (zmanian):
- H1: Accept configurable embedding dimension in LanceDbVectorStore::new()
  instead of hardcoding 1536. Dimension is sourced from
  EmbeddingProvider::dimension() at init time.
- H2: Skip double-write of embeddings to DB when external vector store
  is active (pass None to insert_chunk for embedding column).
- H3: Update PR title from "refactor" to "feat" (net-new feature).
- H4: Document non-atomic update_embedding in struct doc comment.

Bot reviewer (Copilot):
- Cache LanceDB table handle via tokio::sync::OnceCell (avoid
  open_table per operation).
- Cache Arc<Schema> in struct (avoid rebuilding per insert).
- Fix error variants: ChunkingFailed → EmbeddingFailed for LanceDB
  store/delete operations.
- Propagate store_embedding errors in reindex_document instead of
  warn-only (prevents silent data loss).
- Prefetch document metadata map in backfill_embeddings to avoid N+1
  queries.
- Add lancedb feature + protoc to CI test matrix so LanceDB tests
  actually run on Linux.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

.env.example

@@ -2,6 +2,12 @@
 DATABASE_URL=postgres://localhost/ironclaw
 DATABASE_POOL_SIZE=10
 
+# Vector store for workspace memory (optional)
+# When set to "lancedb", uses LanceDB for semantic search instead of pgvector/libsql
+# VECTOR_BACKEND=builtin   # default: use database's built-in index (pgvector or libsql_vector_idx); "pgvector" is also accepted as an alias for "builtin"
+# VECTOR_BACKEND=lancedb
+# LANCEDB_PATH=~/.ironclaw/lancedb   # path for LanceDB when VECTOR_BACKEND=lancedb
+
 # LLM Provider
 # LLM_BACKEND=nearai           # default
 # Possible values: nearai, ollama, openai_compatible, openai, anthropic, tinfoil

.github/workflows/code_style.yml

@@ -36,6 +36,11 @@ jobs:
       uses: dtolnay/rust-toolchain@stable
       with:
         components: clippy
+    - name: Install protoc (for lancedb)
+      if: matrix.name == 'all-features'
+      uses: arduino/setup-protoc@v3
+      with:
+        repo-token: ${{ secrets.GITHUB_TOKEN }}
     - uses: Swatinem/rust-cache@v2
       with:
         key: clippy-${{ matrix.name }}
@@ -62,6 +67,11 @@ jobs:
       uses: dtolnay/rust-toolchain@stable
       with:
         components: clippy
+    - name: Install protoc (for lancedb)
+      if: matrix.name == 'all-features'
+      uses: arduino/setup-protoc@v3
+      with:
+        repo-token: ${{ secrets.GITHUB_TOKEN }}
     - uses: Swatinem/rust-cache@v2
       with:
         key: clippy-windows-${{ matrix.name }}

.github/workflows/test.yml

@@ -14,7 +14,7 @@ jobs:
       matrix:
         include:
           - name: all-features
-            flags: "--features postgres,libsql,html-to-markdown"
+            flags: "--features postgres,libsql,lancedb,html-to-markdown"
           - name: default
             flags: ""
           - name: libsql-only
@@ -26,6 +26,11 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
         with:
           targets: wasm32-wasip2
+      - name: Install protoc (for lancedb)
+        if: matrix.name == 'all-features'
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: ${{ matrix.name }}
@@ -66,6 +71,11 @@ jobs:
         uses: actions/checkout@v6
       - name: Install Rust
         uses: dtolnay/rust-toolchain@stable
+      - name: Install protoc (for lancedb)
+        if: matrix.name == 'all-features'
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: windows-${{ matrix.name }}
@@ -82,6 +92,10 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
         with:
           targets: wasm32-wasip2
+      - name: Install protoc (for lancedb)
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: wasm-extensions

Cargo.toml

@@ -125,6 +125,11 @@ open = "5"
 # The postgres feature provides ToSql/FromSql for postgres-types (shared by tokio-postgres)
 pgvector = { version = "0.4", features = ["postgres"], optional = true }
 
+# LanceDB vector store (optional alternative to pgvector/libsql for workspace search)
+lancedb = { version = "0.26", optional = true }
+arrow-array = { version = "57", optional = true }
+arrow-schema = { version = "57", optional = true }
+
 # WASM sandbox for untrusted tool execution
 wasmtime = { version = "28", features = ["component-model"] }
 wasmtime-wasi = "28"  # WASI support for component model
@@ -189,6 +194,7 @@ insta = "1.46.3"
 
 [features]
 default = ["postgres", "libsql", "html-to-markdown"]
+lancedb = ["dep:lancedb", "dep:arrow-array", "dep:arrow-schema"]
 postgres = [
     "dep:deadpool-postgres",
     "dep:tokio-postgres",

FEATURE_PARITY.md

@@ -340,7 +340,7 @@ This document tracks feature parity between IronClaw (Rust implementation) and O
 | Gemini embeddings | ✅ | ❌ | |
 | Local embeddings | ✅ | ❌ | |
 | SQLite-vec backend | ✅ | ❌ | IronClaw uses PostgreSQL |
-| LanceDB backend | ✅ | ❌ | Configurable auto-capture max length |
+| LanceDB backend | ✅ | ✅ | VectorStore trait + LanceDbVectorStore (configured via VECTOR_BACKEND=lancedb) |
 | QMD backend | ✅ | ❌ | |
 | Atomic reindexing | ✅ | ✅ | |
 | Embeddings batching | ✅ | ✅ | `embed_batch` on EmbeddingProvider trait |

src/app.rs

@@ -386,12 +386,53 @@ impl AppBuilder {
             .embeddings
             .create_provider(&self.config.llm.nearai.base_url, self.session.clone());
 
+        // Create optional external vector store for workspace semantic search
+        let vector_store: Option<Arc<dyn crate::workspace::VectorStore>> = {
+            #[cfg(feature = "lancedb")]
+            {
+                if self.config.database.vector_backend == crate::config::VectorBackend::LanceDb {
+                    let path = self
+                        .config
+                        .database
+                        .lancedb_path
+                        .clone()
+                        .unwrap_or_else(crate::config::default_lancedb_path);
+                    let dim = embeddings.as_ref().map(|p| p.dimension());
+                    match crate::workspace::LanceDbVectorStore::new(path, dim).await {
+                        Ok(store) => {
+                            tracing::info!("LanceDB vector store connected for workspace search");
+                            Some(Arc::new(store) as Arc<dyn crate::workspace::VectorStore>)
+                        }
+                        Err(e) => {
+                            tracing::warn!("Failed to initialize LanceDB: {}", e);
+                            None
+                        }
+                    }
+                } else {
+                    None
+                }
+            }
+            #[cfg(not(feature = "lancedb"))]
+            {
+                if self.config.database.vector_backend == crate::config::VectorBackend::LanceDb {
+                    tracing::warn!(
+                        "VECTOR_BACKEND=lancedb but 'lancedb' feature not enabled; \
+                         falling back to built-in vector search"
+                    );
+                }
+                None
+            }
+        };
+
         // Register memory tools if database is available
         let workspace = if let Some(ref db) = self.db {
             let mut ws = Workspace::new_with_db("default", db.clone());
             if let Some(ref emb) = embeddings {
                 ws = ws.with_embeddings(emb.clone());
             }
+            if let Some(ref vs) = vector_store {
+                ws = ws.with_vector_store(vs.clone());
+            }
             let ws = Arc::new(ws);
             tools.register_memory_tools(Arc::clone(&ws));
             Some(ws)

src/config/database.rs

@@ -82,6 +82,31 @@ impl std::str::FromStr for SslMode {
     }
 }
 
+/// Which vector store backend to use for workspace semantic search.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum VectorBackend {
+    /// Use the database's built-in vector support (pgvector or libsql_vector_idx).
+    #[default]
+    Builtin,
+    /// Use LanceDB as an external vector store (requires `lancedb` feature).
+    LanceDb,
+}
+
+impl std::str::FromStr for VectorBackend {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_lowercase().as_str() {
+            "" | "builtin" | "pgvector" | "libsql" => Ok(Self::Builtin),
+            "lancedb" | "lance" => Ok(Self::LanceDb),
+            _ => Err(format!(
+                "invalid vector backend '{}', expected 'builtin' or 'lancedb'",
+                s
+            )),
+        }
+    }
+}
+
 /// Database configuration.
 #[derive(Debug, Clone)]
 pub struct DatabaseConfig {
@@ -101,6 +126,12 @@ pub struct DatabaseConfig {
     pub libsql_url: Option<String>,
     /// Turso auth token (required when libsql_url is set).
     pub libsql_auth_token: Option<SecretString>,
+
+    // -- Vector store fields --
+    /// Which vector store to use for workspace semantic search (default: Builtin).
+    pub vector_backend: VectorBackend,
+    /// Path to LanceDB directory (default: ~/.ironclaw/lancedb when vector_backend is LanceDb).
+    pub lancedb_path: Option<PathBuf>,
 }
 
 impl DatabaseConfig {
@@ -159,6 +190,25 @@ impl DatabaseConfig {
             });
         }
 
+        let vector_backend: VectorBackend = if let Some(s) = optional_env("VECTOR_BACKEND")? {
+            s.parse().map_err(|e| ConfigError::InvalidValue {
+                key: "VECTOR_BACKEND".to_string(),
+                message: e,
+            })?
+        } else {
+            VectorBackend::default()
+        };
+
+        let lancedb_path = optional_env("LANCEDB_PATH")?
+            .map(PathBuf::from)
+            .or_else(|| {
+                if vector_backend == VectorBackend::LanceDb {
+                    Some(default_lancedb_path())
+                } else {
+                    None
+                }
+            });
+
         Ok(Self {
             backend,
             url: SecretString::from(url),
@@ -167,6 +217,8 @@ impl DatabaseConfig {
             libsql_path,
             libsql_url,
             libsql_auth_token,
+            vector_backend,
+            lancedb_path,
         })
     }
 
@@ -195,6 +247,11 @@ pub fn default_libsql_path() -> PathBuf {
     ironclaw_base_dir().join("ironclaw.db")
 }
 
+/// Default LanceDB directory (~/.ironclaw/lancedb).
+pub fn default_lancedb_path() -> PathBuf {
+    ironclaw_base_dir().join("lancedb")
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -224,4 +281,39 @@ mod tests {
     fn ssl_mode_parse_invalid() {
         assert!("invalid".parse::<SslMode>().is_err());
     }
+
+    #[test]
+    fn vector_backend_parse() {
+        assert_eq!(
+            "builtin".parse::<VectorBackend>().unwrap(),
+            VectorBackend::Builtin
+        );
+        assert_eq!(
+            "pgvector".parse::<VectorBackend>().unwrap(),
+            VectorBackend::Builtin
+        );
+        assert_eq!(
+            "libsql".parse::<VectorBackend>().unwrap(),
+            VectorBackend::Builtin
+        );
+        assert_eq!("".parse::<VectorBackend>().unwrap(), VectorBackend::Builtin);
+
+        assert_eq!(
+            "lancedb".parse::<VectorBackend>().unwrap(),
+            VectorBackend::LanceDb
+        );
+        assert_eq!(
+            "lance".parse::<VectorBackend>().unwrap(),
+            VectorBackend::LanceDb
+        );
+
+        assert!("invalid".parse::<VectorBackend>().is_err());
+    }
+
+    #[test]
+    fn default_lancedb_path_under_ironclaw() {
+        let path = super::default_lancedb_path();
+        assert!(path.to_string_lossy().contains("ironclaw"));
+        assert!(path.to_string_lossy().ends_with("lancedb"));
+    }
 }

src/config/mod.rs

@@ -33,7 +33,10 @@ use crate::settings::Settings;
 pub use self::agent::AgentConfig;
 pub use self::builder::BuilderModeConfig;
 pub use self::channels::{ChannelsConfig, CliConfig, GatewayConfig, HttpConfig, SignalConfig};
-pub use self::database::{DatabaseBackend, DatabaseConfig, SslMode, default_libsql_path};
+pub use self::database::{
+    DatabaseBackend, DatabaseConfig, SslMode, VectorBackend, default_lancedb_path,
+    default_libsql_path,
+};
 pub use self::embeddings::EmbeddingsConfig;
 pub use self::heartbeat::HeartbeatConfig;
 pub use self::hygiene::HygieneConfig;
@@ -107,6 +110,8 @@ impl Config {
                 libsql_path: Some(libsql_path),
                 libsql_url: None,
                 libsql_auth_token: None,
+                vector_backend: VectorBackend::default(),
+                lancedb_path: None,
             },
             llm: LlmConfig::for_testing(),
             embeddings: EmbeddingsConfig::default(),

src/workspace/lancedb_store.rs

@@ -0,0 +1,644 @@
+//! LanceDB-backed vector store for workspace memory chunks.
+//!
+//! Provides an alternative to pgvector/libsql for semantic search when the
+//! `lancedb` feature is enabled. Documents and metadata stay in the main
+//! database; this store holds chunk embeddings for vector similarity search.
+//!
+//! Configuration:
+//!   LANCEDB_PATH=~/.ironclaw/lancedb   # Default
+//!   VECTOR_BACKEND=lancedb             # Use LanceDB for vector search
+
+/// Default embedding dimension (text-embedding-3-small).
+/// Override by passing the actual provider dimension to `LanceDbVectorStore::new()`.
+pub const DEFAULT_EMBEDDING_DIM: i32 = 1536;
+
+#[cfg(feature = "lancedb")]
+mod impl_lancedb {
+    use std::sync::Arc;
+
+    use arrow_array::types::Float32Type;
+    use arrow_array::{Array, FixedSizeListArray, RecordBatch, RecordBatchIterator, StringArray};
+    use arrow_schema::{DataType, Field, Schema};
+    use async_trait::async_trait;
+    use futures::StreamExt;
+    use lancedb::query::{ExecutableQuery, QueryBase};
+    use uuid::Uuid;
+
+    use super::DEFAULT_EMBEDDING_DIM;
+    use crate::error::WorkspaceError;
+    use crate::workspace::search::RankedResult;
+    use crate::workspace::vector_store::VectorStore;
+
+    const TABLE_NAME: &str = "memory_chunks";
+
+    /// Escapes a string for safe use in LanceDB predicate expressions.
+    /// Uses SQL-style escaping: single quotes are doubled to prevent injection.
+    fn escape_predicate_value(s: &str) -> String {
+        s.replace('\'', "''")
+    }
+
+    /// LanceDB-backed vector store.
+    ///
+    /// The `update_embedding` method uses delete-then-insert (not atomic).
+    /// LanceDB does not support transactions, so a crash between the two
+    /// operations can lose the embedding for that chunk. This is acceptable
+    /// for personal workspace sizes where data can be reindexed.
+    pub struct LanceDbVectorStore {
+        db: Arc<lancedb::Connection>,
+        table_name: String,
+        embedding_dim: i32,
+        schema: Arc<Schema>,
+        table: tokio::sync::OnceCell<lancedb::Table>,
+    }
+
+    impl LanceDbVectorStore {
+        /// Create a new LanceDB store at the given path.
+        ///
+        /// `embedding_dim` should match `EmbeddingProvider::dimension()`.
+        /// Pass `None` to use the default (1536, text-embedding-3-small).
+        pub async fn new(
+            path: impl AsRef<std::path::Path>,
+            embedding_dim: Option<usize>,
+        ) -> Result<Self, WorkspaceError> {
+            let path_str = path
+                .as_ref()
+                .to_str()
+                .ok_or_else(|| WorkspaceError::SearchFailed {
+                    reason: "Invalid LanceDB path".to_string(),
+                })?;
+
+            let db = lancedb::connect(path_str).execute().await.map_err(|e| {
+                WorkspaceError::SearchFailed {
+                    reason: format!("Failed to connect to LanceDB: {}", e),
+                }
+            })?;
+
+            let dim = embedding_dim.unwrap_or(DEFAULT_EMBEDDING_DIM as usize) as i32;
+            let schema = Arc::new(Self::build_schema(dim));
+
+            let store = Self {
+                db: Arc::new(db),
+                table_name: TABLE_NAME.to_string(),
+                embedding_dim: dim,
+                schema,
+                table: tokio::sync::OnceCell::new(),
+            };
+
+            store.ensure_table().await?;
+            Ok(store)
+        }
+
+        async fn ensure_table(&self) -> Result<(), WorkspaceError> {
+            let tables = self.db.table_names().execute().await.map_err(|e| {
+                WorkspaceError::SearchFailed {
+                    reason: format!("Failed to list tables: {}", e),
+                }
+            })?;
+
+            if tables.iter().any(|t| t == &self.table_name) {
+                return Ok(());
+            }
+
+            self.db
+                .create_empty_table(&self.table_name, self.schema.clone())
+                .execute()
+                .await
+                .map_err(|e| WorkspaceError::SearchFailed {
+                    reason: format!("Failed to create table: {}", e),
+                })?;
+
+            // Index creation is deferred — brute-force search via
+            // bypass_vector_index() works without a pre-built index and is
+            // sufficient for personal workspace sizes.
+            Ok(())
+        }
+
+        /// Get or open the cached table handle.
+        async fn table(&self) -> Result<&lancedb::Table, WorkspaceError> {
+            self.table
+                .get_or_try_init(|| async {
+                    self.db
+                        .open_table(&self.table_name)
+                        .execute()
+                        .await
+                        .map_err(|e| WorkspaceError::SearchFailed {
+                            reason: format!("Failed to open table: {}", e),
+                        })
+                })
+                .await
+        }
+
+        fn build_schema(embedding_dim: i32) -> Schema {
+            Schema::new(vec![
+                Field::new("chunk_id", DataType::Utf8, false),
+                Field::new("document_id", DataType::Utf8, false),
+                Field::new("document_path", DataType::Utf8, false),
+                Field::new("user_id", DataType::Utf8, false),
+                Field::new("agent_id", DataType::Utf8, true),
+                Field::new("content", DataType::Utf8, false),
+                Field::new(
+                    "vector",
+                    DataType::FixedSizeList(
+                        Arc::new(Field::new("item", DataType::Float32, true)),
+                        embedding_dim,
+                    ),
+                    false,
+                ),
+            ])
+        }
+    }
+
+    #[async_trait]
+    impl VectorStore for LanceDbVectorStore {
+        async fn store_embedding(
+            &self,
+            chunk_id: Uuid,
+            document_id: Uuid,
+            document_path: &str,
+            user_id: &str,
+            agent_id: Option<Uuid>,
+            content: &str,
+            embedding: &[f32],
+        ) -> Result<(), WorkspaceError> {
+            if embedding.len() != self.embedding_dim as usize {
+                return Err(WorkspaceError::EmbeddingFailed {
+                    reason: format!(
+                        "Embedding dimension {} does not match expected {}",
+                        embedding.len(),
+                        self.embedding_dim
+                    ),
+                });
+            }
+
+            let table = self.table().await?;
+
+            let chunk_ids = StringArray::from(vec![chunk_id.to_string()]);
+            let document_ids = StringArray::from(vec![document_id.to_string()]);
+            let document_paths = StringArray::from(vec![document_path]);
+            let user_ids = StringArray::from(vec![user_id]);
+            let agent_ids = StringArray::from(vec![agent_id.map(|a| a.to_string())]);
+            let contents = StringArray::from(vec![content]);
+            let vec_values: Vec<Option<f32>> = embedding.iter().map(|&x| Some(x)).collect();
+            let vectors = FixedSizeListArray::from_iter_primitive::<Float32Type, _, _>(
+                vec![Some(vec_values)],
+                self.embedding_dim,
+            );
+
+            let batch = RecordBatch::try_new(
+                self.schema.clone(),
+                vec![
+                    Arc::new(chunk_ids),
+                    Arc::new(document_ids),
+                    Arc::new(document_paths),
+                    Arc::new(user_ids),
+                    Arc::new(agent_ids),
+                    Arc::new(contents),
+                    Arc::new(vectors),
+                ],
+            )
+            .map_err(|e| WorkspaceError::EmbeddingFailed {
+                reason: format!("Failed to create record batch: {}", e),
+            })?;
+
+            let batches =
+                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), self.schema.clone());
+
+            table
+                .add(Box::new(batches) as Box<dyn arrow_array::RecordBatchReader + Send>)
+                .execute()
+                .await
+                .map_err(|e| WorkspaceError::EmbeddingFailed {
+                    reason: format!("Failed to store embedding: {}", e),
+                })?;
+
+            Ok(())
+        }
+
+        async fn update_embedding(
+            &self,
+            chunk_id: Uuid,
+            document_id: Uuid,
+            document_path: &str,
+            user_id: &str,
+            agent_id: Option<Uuid>,
+            content: &str,
+            embedding: &[f32],
+        ) -> Result<(), WorkspaceError> {
+            let table = self.table().await?;
+
+            table
+                .delete(&format!(
+                    "chunk_id = '{}'",
+                    escape_predicate_value(&chunk_id.to_string())
+                ))
+                .await
+                .map_err(|e| WorkspaceError::EmbeddingFailed {
+                    reason: format!("Failed to delete chunk for update: {}", e),
+                })?;
+
+            self.store_embedding(
+                chunk_id,
+                document_id,
+                document_path,
+                user_id,
+                agent_id,
+                content,
+                embedding,
+            )
+            .await
+        }
+
+        async fn delete_embeddings(&self, document_id: Uuid) -> Result<(), WorkspaceError> {
+            let table = self.table().await?;
+
+            table
+                .delete(&format!(
+                    "document_id = '{}'",
+                    escape_predicate_value(&document_id.to_string())
+                ))
+                .await
+                .map_err(|e| WorkspaceError::EmbeddingFailed {
+                    reason: format!("Failed to delete embeddings: {}", e),
+                })?;
+
+            Ok(())
+        }
+
+        async fn vector_search(
+            &self,
+            user_id: &str,
+            agent_id: Option<Uuid>,
+            embedding: &[f32],
+            limit: usize,
+        ) -> Result<Vec<RankedResult>, WorkspaceError> {
+            let table = self.table().await?;
+
+            let filter = if let Some(aid) = agent_id {
+                format!(
+                    "user_id = '{}' AND agent_id = '{}'",
+                    escape_predicate_value(user_id),
+                    escape_predicate_value(&aid.to_string())
+                )
+            } else {
+                format!(
+                    "user_id = '{}' AND agent_id IS NULL",
+                    escape_predicate_value(user_id)
+                )
+            };
+
+            let query = table
+                .query()
+                .nearest_to(embedding)
+                .map_err(|e| WorkspaceError::SearchFailed {
+                    reason: format!("Invalid query vector: {}", e),
+                })?
+                .only_if(&filter)
+                .bypass_vector_index()
+                .limit(limit);
+            let mut stream = ExecutableQuery::execute(&query).await.map_err(|e| {
+                WorkspaceError::SearchFailed {
+                    reason: format!("Vector search failed: {}", e),
+                }
+            })?;
+
+            let mut results = Vec::new();
+            let mut rank: u32 = 1;
+            while let Some(batch_result) = stream.next().await {
+                let batch = batch_result.map_err(|e| WorkspaceError::SearchFailed {
+                    reason: format!("Stream error: {}", e),
+                })?;
+
+                let chunk_id_col = batch.column_by_name("chunk_id").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "chunk_id column missing".to_string(),
+                    }
+                })?;
+                let document_id_col = batch.column_by_name("document_id").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "document_id column missing".to_string(),
+                    }
+                })?;
+                let document_path_col = batch.column_by_name("document_path").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "document_path column missing".to_string(),
+                    }
+                })?;
+                let content_col = batch.column_by_name("content").ok_or_else(|| {
+                    WorkspaceError::SearchFailed {
+                        reason: "content column missing".to_string(),
+                    }
+                })?;
+
+                let chunk_ids = chunk_id_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "chunk_id wrong type".to_string(),
+                    })?;
+                let document_ids = document_id_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "document_id wrong type".to_string(),
+                    })?;
+                let document_paths = document_path_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "document_path wrong type".to_string(),
+                    })?;
+                let contents = content_col
+                    .as_any()
+                    .downcast_ref::<StringArray>()
+                    .ok_or_else(|| WorkspaceError::SearchFailed {
+                        reason: "content wrong type".to_string(),
+                    })?;
+
+                for i in 0..batch.num_rows() {
+                    let raw_chunk_id = chunk_ids.value(i);
+                    let chunk_id =
+                        raw_chunk_id
+                            .parse::<Uuid>()
+                            .map_err(|e| WorkspaceError::SearchFailed {
+                                reason: format!("Invalid chunk_id UUID '{}': {}", raw_chunk_id, e),
+                            })?;
+
+                    let raw_document_id = document_ids.value(i);
+                    let document_id = raw_document_id.parse::<Uuid>().map_err(|e| {
+                        WorkspaceError::SearchFailed {
+                            reason: format!(
+                                "Invalid document_id UUID '{}': {}",
+                                raw_document_id, e
+                            ),
+                        }
+                    })?;
+                    let document_path = document_paths.value(i).to_string();
+                    let content = contents.value(i).to_string();
+
+                    results.push(RankedResult {
+                        chunk_id,
+                        document_id,
+                        document_path,
+                        content,
+                        rank,
+                    });
+                    rank += 1;
+                }
+            }
+
+            Ok(results)
+        }
+    }
+}
+
+#[cfg(feature = "lancedb")]
+pub use impl_lancedb::LanceDbVectorStore;
+
+#[cfg(all(test, feature = "lancedb"))]
+mod tests {
+    use tempfile::TempDir;
+    use uuid::Uuid;
+
+    use super::{DEFAULT_EMBEDDING_DIM, LanceDbVectorStore};
+    use crate::workspace::vector_store::VectorStore;
+
+    fn make_embedding(seed: f32) -> Vec<f32> {
+        (0..DEFAULT_EMBEDDING_DIM as usize)
+            .map(|i| (seed * (i as f32 + 1.0)).sin())
+            .collect()
+    }
+
+    #[tokio::test]
+    async fn test_insert_and_vector_search() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let chunk_id = Uuid::new_v4();
+        let document_id = Uuid::new_v4();
+        let user_id = "user1";
+        let content = "Rust is a systems programming language";
+        let embedding = make_embedding(1.0);
+
+        store
+            .store_embedding(
+                chunk_id,
+                document_id,
+                "test.md",
+                user_id,
+                None,
+                content,
+                &embedding,
+            )
+            .await
+            .unwrap();
+
+        let results = store
+            .vector_search(user_id, None, &embedding, 5)
+            .await
+            .unwrap();
+
+        assert_eq!(results.len(), 1);
+        assert_eq!(results[0].chunk_id, chunk_id);
+        assert_eq!(results[0].document_id, document_id);
+        assert_eq!(results[0].content, content);
+        assert_eq!(results[0].rank, 1);
+    }
+
+    #[tokio::test]
+    async fn test_insert_multiple_and_search_returns_ordered() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let doc_id = Uuid::new_v4();
+        let user_id = "user1";
+
+        for (i, seed) in [1.0, 2.0, 3.0].iter().enumerate() {
+            store
+                .store_embedding(
+                    Uuid::new_v4(),
+                    doc_id,
+                    "test.md",
+                    user_id,
+                    None,
+                    &format!("content {}", i),
+                    &make_embedding(*seed),
+                )
+                .await
+                .unwrap();
+        }
+
+        let query_emb = make_embedding(2.0);
+        let results = store
+            .vector_search(user_id, None, &query_emb, 5)
+            .await
+            .unwrap();
+
+        assert_eq!(results.len(), 3);
+        let contents: Vec<_> = results.iter().map(|r| r.content.as_str()).collect();
+        assert!(contents.contains(&"content 0"));
+        assert!(contents.contains(&"content 1"));
+        assert!(contents.contains(&"content 2"));
+    }
+
+    #[tokio::test]
+    async fn test_delete_chunks() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let doc_id = Uuid::new_v4();
+        let user_id = "user1";
+
+        store
+            .store_embedding(
+                Uuid::new_v4(),
+                doc_id,
+                "test.md",
+                user_id,
+                None,
+                "content",
+                &make_embedding(1.0),
+            )
+            .await
+            .unwrap();
+
+        let results = store
+            .vector_search(user_id, None, &make_embedding(1.0), 5)
+            .await
+            .unwrap();
+        assert_eq!(results.len(), 1);
+
+        store.delete_embeddings(doc_id).await.unwrap();
+
+        let results_after = store
+            .vector_search(user_id, None, &make_embedding(1.0), 5)
+            .await
+            .unwrap();
+        assert!(results_after.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_update_chunk_embedding() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let chunk_id = Uuid::new_v4();
+        let doc_id = Uuid::new_v4();
+        let user_id = "user1";
+        let content = "original content";
+
+        store
+            .store_embedding(
+                chunk_id,
+                doc_id,
+                "test.md",
+                user_id,
+                None,
+                content,
+                &make_embedding(1.0),
+            )
+            .await
+            .unwrap();
+
+        let new_embedding = make_embedding(5.0);
+        store
+            .update_embedding(
+                chunk_id,
+                doc_id,
+                "test.md",
+                user_id,
+                None,
+                content,
+                &new_embedding,
+            )
+            .await
+            .unwrap();
+
+        let results = store
+            .vector_search(user_id, None, &new_embedding, 5)
+            .await
+            .unwrap();
+        assert_eq!(results.len(), 1);
+        assert_eq!(results[0].chunk_id, chunk_id);
+    }
+
+    #[tokio::test]
+    async fn test_vector_search_filters_by_user_and_agent() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let doc_id = Uuid::new_v4();
+        let embedding = make_embedding(1.0);
+
+        store
+            .store_embedding(
+                Uuid::new_v4(),
+                doc_id,
+                "test.md",
+                "user1",
+                None,
+                "user1 content",
+                &embedding,
+            )
+            .await
+            .unwrap();
+
+        store
+            .store_embedding(
+                Uuid::new_v4(),
+                doc_id,
+                "test.md",
+                "user2",
+                None,
+                "user2 content",
+                &embedding,
+            )
+            .await
+            .unwrap();
+
+        let results_user1 = store
+            .vector_search("user1", None, &embedding, 5)
+            .await
+            .unwrap();
+        assert_eq!(results_user1.len(), 1);
+        assert_eq!(results_user1[0].content, "user1 content");
+
+        let results_user2 = store
+            .vector_search("user2", None, &embedding, 5)
+            .await
+            .unwrap();
+        assert_eq!(results_user2.len(), 1);
+        assert_eq!(results_user2[0].content, "user2 content");
+
+        let results_wrong_user = store
+            .vector_search("user3", None, &embedding, 5)
+            .await
+            .unwrap();
+        assert!(results_wrong_user.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_insert_rejects_wrong_embedding_dim() {
+        let dir = TempDir::new().unwrap();
+        let store = LanceDbVectorStore::new(dir.path(), None).await.unwrap();
+
+        let wrong_dim: Vec<f32> = vec![1.0; 100];
+
+        let err = store
+            .store_embedding(
+                Uuid::new_v4(),
+                Uuid::new_v4(),
+                "test.md",
+                "user1",
+                None,
+                "content",
+                &wrong_dim,
+            )
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            err,
+            crate::error::WorkspaceError::EmbeddingFailed { .. }
+        ));
+    }
+}

src/workspace/mod.rs

@@ -42,20 +42,26 @@
 
 mod chunker;
 mod document;
-mod embeddings;
+pub mod embeddings;
 pub mod hygiene;
+#[cfg(feature = "lancedb")]
+pub mod lancedb_store;
 #[cfg(feature = "postgres")]
 mod repository;
 mod search;
+pub mod vector_store;
 
 pub use chunker::{ChunkConfig, chunk_document};
 pub use document::{MemoryChunk, MemoryDocument, WorkspaceEntry, paths};
 pub use embeddings::{
     EmbeddingProvider, MockEmbeddings, NearAiEmbeddings, OllamaEmbeddings, OpenAiEmbeddings,
 };
+#[cfg(feature = "lancedb")]
+pub use lancedb_store::{DEFAULT_EMBEDDING_DIM, LanceDbVectorStore};
 #[cfg(feature = "postgres")]
 pub use repository::Repository;
 pub use search::{RankedResult, SearchConfig, SearchResult, reciprocal_rank_fusion};
+pub use vector_store::VectorStore;
 
 use std::sync::Arc;
 
@@ -332,6 +338,12 @@ pub struct Workspace {
     storage: WorkspaceStorage,
     /// Embedding provider for semantic search.
     embeddings: Option<Arc<dyn EmbeddingProvider>>,
+    /// Optional external vector store for semantic search.
+    ///
+    /// When set, embeddings are stored here instead of (or in addition to)
+    /// the database's built-in vector support, and hybrid search uses this
+    /// for the vector component while FTS comes from the database.
+    vector_store: Option<Arc<dyn VectorStore>>,
 }
 
 impl Workspace {
@@ -343,6 +355,7 @@ impl Workspace {
             agent_id: None,
             storage: WorkspaceStorage::Repo(Repository::new(pool)),
             embeddings: None,
+            vector_store: None,
         }
     }
 
@@ -355,6 +368,7 @@ impl Workspace {
             agent_id: None,
             storage: WorkspaceStorage::Db(db),
             embeddings: None,
+            vector_store: None,
         }
     }
 
@@ -370,6 +384,17 @@ impl Workspace {
         self
     }
 
+    /// Set an external vector store for semantic search.
+    ///
+    /// When set, vector operations (store/search/delete embeddings) use this
+    /// store instead of the database's built-in vector support. FTS continues
+    /// to use the database. Hybrid search combines FTS from the database with
+    /// vector results from this store via RRF.
+    pub fn with_vector_store(mut self, store: Arc<dyn VectorStore>) -> Self {
+        self.vector_store = Some(store);
+        self
+    }
+
     /// Get the user ID.
     pub fn user_id(&self) -> &str {
         &self.user_id
@@ -458,9 +483,21 @@ impl Workspace {
 
     /// Delete a file.
     ///
-    /// Also deletes associated chunks.
+    /// Also deletes associated chunks (from both DB and external vector store).
     pub async fn delete(&self, path: &str) -> Result<(), WorkspaceError> {
         let path = normalize_path(path);
+
+        // Clean up external vector store before DB cascade deletes chunks
+        if let Some(ref vs) = self.vector_store
+            && let Ok(doc) = self
+                .storage
+                .get_document_by_path(&self.user_id, self.agent_id, &path)
+                .await
+            && let Err(e) = vs.delete_embeddings(doc.id).await
+        {
+            tracing::warn!("Failed to delete embeddings from vector store: {}", e);
+        }
+
         self.storage
             .delete_document_by_path(&self.user_id, self.agent_id, &path)
             .await
@@ -725,20 +762,67 @@ impl Workspace {
         query: &str,
         config: SearchConfig,
     ) -> Result<Vec<SearchResult>, WorkspaceError> {
-        // Generate embedding for semantic search if provider available
-        let embedding = if let Some(ref provider) = self.embeddings {
-            Some(
-                provider
-                    .embed(query)
-                    .await
-                    .map_err(|e| WorkspaceError::EmbeddingFailed {
-                        reason: e.to_string(),
-                    })?,
-            )
-        } else {
-            None
-        };
+        // Generate embedding for semantic search only when vector search is enabled
+        let embedding =
+            if config.use_vector {
+                if let Some(ref provider) = self.embeddings {
+                    Some(provider.embed(query).await.map_err(|e| {
+                        WorkspaceError::EmbeddingFailed {
+                            reason: e.to_string(),
+                        }
+                    })?)
+                } else {
+                    None
+                }
+            } else {
+                None
+            };
 
+        // When an external vector store is configured, do FTS from the
+        // database and vector search from the store, then fuse with RRF.
+        if let Some(ref vs) = self.vector_store {
+            // FTS from database (disable vector to avoid double-searching)
+            let fts_results = if config.use_fts {
+                let fts_config = SearchConfig {
+                    use_fts: true,
+                    use_vector: false,
+                    ..config.clone()
+                };
+                let fts_search = self
+                    .storage
+                    .hybrid_search(&self.user_id, self.agent_id, query, None, &fts_config)
+                    .await?;
+                fts_search
+                    .into_iter()
+                    .enumerate()
+                    .map(|(i, r)| RankedResult {
+                        chunk_id: r.chunk_id,
+                        document_id: r.document_id,
+                        document_path: r.document_path,
+                        content: r.content,
+                        rank: (i + 1) as u32,
+                    })
+                    .collect()
+            } else {
+                Vec::new()
+            };
+
+            // Vector search from external store
+            let vector_results = if config.use_vector {
+                if let Some(ref emb) = embedding {
+                    vs.vector_search(&self.user_id, self.agent_id, emb, config.pre_fusion_limit)
+                        .await?
+                } else {
+                    Vec::new()
+                }
+            } else {
+                Vec::new()
+            };
+
+            return Ok(reciprocal_rank_fusion(fts_results, vector_results, &config));
+        }
+
+        // No external vector store — use database's built-in hybrid search
         self.storage
             .hybrid_search(
                 &self.user_id,
@@ -760,7 +844,13 @@ impl Workspace {
         // Chunk the content
         let chunks = chunk_document(&doc.content, ChunkConfig::default());
 
-        // Delete old chunks
+        // Delete old embeddings from external vector store FIRST — if this fails,
+        // we abort before touching DB chunks, keeping the document consistent.
+        if let Some(ref vs) = self.vector_store {
+            vs.delete_embeddings(document_id).await?;
+        }
+
+        // Delete old chunks from database
         self.storage.delete_chunks(document_id).await?;
 
         // Insert new chunks
@@ -778,9 +868,33 @@ impl Workspace {
                 None
             };
 
-            self.storage
-                .insert_chunk(document_id, index as i32, &content, embedding.as_deref())
+            // When an external vector store is active, skip writing embeddings
+            // to the DB (they'd never be queried from there).
+            let db_embedding = if self.vector_store.is_some() {
+                None
+            } else {
+                embedding.as_deref()
+            };
+
+            let chunk_id = self
+                .storage
+                .insert_chunk(document_id, index as i32, &content, db_embedding)
                 .await?;
+
+            // Sync embedding to external vector store (propagate errors to
+            // avoid leaving a document with deleted-then-missing embeddings).
+            if let (Some(vs), Some(emb)) = (&self.vector_store, &embedding) {
+                vs.store_embedding(
+                    chunk_id,
+                    document_id,
+                    &doc.path,
+                    &doc.user_id,
+                    doc.agent_id,
+                    &content,
+                    emb,
+                )
+                .await?;
+            }
         }
 
         Ok(())
@@ -1025,6 +1139,23 @@ impl Workspace {
             .get_chunks_without_embeddings(&self.user_id, self.agent_id, 100)
             .await?;
 
+        // Prefetch document metadata to avoid N+1 queries when syncing to vector store
+        let doc_map: std::collections::HashMap<Uuid, crate::workspace::document::MemoryDocument> =
+            if self.vector_store.is_some() {
+                let mut map = std::collections::HashMap::new();
+                for chunk in &chunks {
+                    if !map.contains_key(&chunk.document_id)
+                        && let Ok(doc) =
+                            self.storage.get_document_by_id(chunk.document_id).await
+                    {
+                        map.insert(doc.id, doc);
+                    }
+                }
+                map
+            } else {
+                std::collections::HashMap::new()
+            };
+
         let mut count = 0;
         for chunk in chunks {
             match provider.embed(&chunk.content).await {
@@ -1032,6 +1163,29 @@ impl Workspace {
                     self.storage
                         .update_chunk_embedding(chunk.id, &embedding)
                         .await?;
+
+                    // Sync to external vector store
+                    if let Some(ref vs) = self.vector_store
+                        && let Some(doc) = doc_map.get(&chunk.document_id)
+                        && let Err(e) = vs
+                            .update_embedding(
+                                chunk.id,
+                                chunk.document_id,
+                                &doc.path,
+                                &doc.user_id,
+                                doc.agent_id,
+                                &chunk.content,
+                                &embedding,
+                            )
+                            .await
+                    {
+                        tracing::warn!(
+                            "Failed to sync embedding to vector store for chunk {}: {}",
+                            chunk.id,
+                            e
+                        );
+                    }
+
                     count += 1;
                 }
                 Err(e) => {

src/workspace/vector_store.rs

@@ -0,0 +1,71 @@
+//! Vector store abstraction for workspace semantic search.
+//!
+//! Separates vector search from the main `Database` trait so that
+//! third-party vector backends (LanceDB, Qdrant, Pinecone, etc.) can
+//! be added by implementing a 4-method trait instead of wrapping the
+//! entire ~80-method `Database` trait.
+//!
+//! When no external vector store is configured, the built-in database
+//! vector support (pgvector / libsql_vector_idx) is used via the
+//! `Database::hybrid_search` method directly.
+
+use async_trait::async_trait;
+use uuid::Uuid;
+
+use crate::error::WorkspaceError;
+use crate::workspace::search::RankedResult;
+
+/// External vector store for semantic search.
+///
+/// Implementations hold chunk embeddings and perform vector similarity
+/// queries. Document/chunk metadata and FTS stay in the main database;
+/// only embeddings live here.
+///
+/// # Adding a new backend
+///
+/// 1. Implement this trait for your backend (4 methods).
+/// 2. Feature-gate the module (`#[cfg(feature = "mybackend")]`).
+/// 3. Pass `Arc<dyn VectorStore>` to `Workspace::with_vector_store()`.
+///
+/// That's it — no Database wrapper, no delegation boilerplate.
+#[async_trait]
+#[allow(clippy::too_many_arguments)]
+pub trait VectorStore: Send + Sync {
+    /// Store an embedding for a chunk.
+    async fn store_embedding(
+        &self,
+        chunk_id: Uuid,
+        document_id: Uuid,
+        document_path: &str,
+        user_id: &str,
+        agent_id: Option<Uuid>,
+        content: &str,
+        embedding: &[f32],
+    ) -> Result<(), WorkspaceError>;
+
+    /// Update an existing chunk's embedding (delete + re-insert is fine).
+    async fn update_embedding(
+        &self,
+        chunk_id: Uuid,
+        document_id: Uuid,
+        document_path: &str,
+        user_id: &str,
+        agent_id: Option<Uuid>,
+        content: &str,
+        embedding: &[f32],
+    ) -> Result<(), WorkspaceError>;
+
+    /// Delete all embeddings for a document.
+    async fn delete_embeddings(&self, document_id: Uuid) -> Result<(), WorkspaceError>;
+
+    /// Vector similarity search, filtered by user and optional agent.
+    ///
+    /// Returns results ranked by similarity (rank 1 = most similar).
+    async fn vector_search(
+        &self,
+        user_id: &str,
+        agent_id: Option<Uuid>,
+        embedding: &[f32],
+        limit: usize,
+    ) -> Result<Vec<RankedResult>, WorkspaceError>;
+}

tests/lancedb_integration.rs

@@ -0,0 +1,123 @@
+//! Integration tests for LanceDB vector store with Workspace composition.
+//!
+//! Requires: cargo test --features "libsql,lancedb"
+//!
+//! Verifies that Workspace correctly composes FTS from libSQL with vector
+//! search from LanceDB via the VectorStore trait.
+
+#![cfg(all(feature = "libsql", feature = "lancedb"))]
+
+use std::sync::Arc;
+
+use ironclaw::db::Database;
+use ironclaw::db::libsql::LibSqlBackend;
+use ironclaw::workspace::{LanceDbVectorStore, SearchConfig, Workspace};
+use tempfile::TempDir;
+
+const EMBEDDING_DIM: usize = 1536;
+
+fn make_embedding(seed: f32) -> Vec<f32> {
+    (0..EMBEDDING_DIM)
+        .map(|i| (seed * (i as f32 + 1.0)).sin())
+        .collect()
+}
+
+/// Mock embedding provider that returns deterministic embeddings.
+struct FixedEmbeddings {
+    embedding: Vec<f32>,
+}
+
+#[async_trait::async_trait]
+impl ironclaw::workspace::EmbeddingProvider for FixedEmbeddings {
+    fn dimension(&self) -> usize {
+        EMBEDDING_DIM
+    }
+
+    fn model_name(&self) -> &str {
+        "fixed-test"
+    }
+
+    fn max_input_length(&self) -> usize {
+        8192
+    }
+
+    async fn embed(
+        &self,
+        _text: &str,
+    ) -> Result<Vec<f32>, ironclaw::workspace::embeddings::EmbeddingError> {
+        Ok(self.embedding.clone())
+    }
+}
+
+async fn setup_workspace() -> (Workspace, TempDir, TempDir) {
+    // Use a temp file (not :memory:) because libSQL in-memory DBs are connection-local
+    let db_dir = TempDir::new().unwrap();
+    let db_path = db_dir.path().join("test.db");
+    let libsql = LibSqlBackend::new_local(&db_path).await.unwrap();
+    libsql.run_migrations().await.unwrap();
+
+    let lancedb_dir = TempDir::new().unwrap();
+    let store = LanceDbVectorStore::new(lancedb_dir.path(), None).await.unwrap();
+
+    let embedding = make_embedding(1.0);
+    let ws = Workspace::new_with_db("test_user", Arc::new(libsql) as Arc<dyn Database>)
+        .with_vector_store(Arc::new(store))
+        .with_embeddings(Arc::new(FixedEmbeddings { embedding }));
+
+    (ws, lancedb_dir, db_dir)
+}
+
+#[tokio::test]
+async fn test_workspace_hybrid_search_with_lancedb() {
+    let (ws, _keep_lance, _keep_db) = setup_workspace().await;
+
+    // Write a document — this triggers chunking + embedding + LanceDB sync
+    ws.write(
+        "context/rust.md",
+        "Rust is a systems programming language focused on safety.",
+    )
+    .await
+    .unwrap();
+
+    // Hybrid search: FTS for "Rust" + vector from LanceDB
+    let results = ws.search("Rust", 5).await.unwrap();
+
+    assert!(!results.is_empty(), "hybrid search should return results");
+    assert!(results[0].content.contains("Rust"));
+}
+
+#[tokio::test]
+async fn test_workspace_delete_removes_from_lancedb() {
+    let (ws, _keep_lance, _keep_db) = setup_workspace().await;
+
+    ws.write("notes/deleted.md", "Content to be deleted.")
+        .await
+        .unwrap();
+
+    let before = ws.search("deleted", 5).await.unwrap();
+    assert_eq!(before.len(), 1);
+
+    ws.delete("notes/deleted.md").await.unwrap();
+
+    let after = ws.search("deleted", 5).await.unwrap();
+    assert!(after.is_empty());
+}
+
+#[tokio::test]
+async fn test_workspace_vector_only_search_uses_lancedb() {
+    let (ws, _keep_lance, _keep_db) = setup_workspace().await;
+
+    ws.write("sync/test.md", "Semantic content for vector search")
+        .await
+        .unwrap();
+
+    // Vector-only search should find via LanceDB even with non-matching FTS query
+    let config = SearchConfig::default().vector_only().with_limit(5);
+    let results = ws
+        .search_with_config("nonexistent_fts_term", config)
+        .await
+        .unwrap();
+
+    assert_eq!(results.len(), 1);
+    assert!(results[0].content.contains("Semantic content"));
+}