1use crate::{parsing::Document, SEMANTIC_INDEX_VERSION};
2use anyhow::{anyhow, Context, Result};
3use project::Fs;
4use rpc::proto::Timestamp;
5use rusqlite::{
6 params,
7 types::{FromSql, FromSqlResult, ValueRef},
8};
9use std::{
10 cmp::Ordering,
11 collections::HashMap,
12 ops::Range,
13 path::{Path, PathBuf},
14 rc::Rc,
15 sync::Arc,
16 time::SystemTime,
17};
18
19#[derive(Debug)]
20pub struct FileRecord {
21 pub id: usize,
22 pub relative_path: String,
23 pub mtime: Timestamp,
24}
25
26#[derive(Debug)]
27struct Embedding(pub Vec<f32>);
28
29impl FromSql for Embedding {
30 fn column_result(value: ValueRef) -> FromSqlResult<Self> {
31 let bytes = value.as_blob()?;
32 let embedding: Result<Vec<f32>, Box<bincode::ErrorKind>> = bincode::deserialize(bytes);
33 if embedding.is_err() {
34 return Err(rusqlite::types::FromSqlError::Other(embedding.unwrap_err()));
35 }
36 return Ok(Embedding(embedding.unwrap()));
37 }
38}
39
40pub struct VectorDatabase {
41 db: rusqlite::Connection,
42}
43
44impl VectorDatabase {
45 pub async fn new(fs: Arc<dyn Fs>, path: Arc<PathBuf>) -> Result<Self> {
46 if let Some(db_directory) = path.parent() {
47 fs.create_dir(db_directory).await?;
48 }
49
50 let this = Self {
51 db: rusqlite::Connection::open(path.as_path())?,
52 };
53 this.initialize_database()?;
54 Ok(this)
55 }
56
57 fn get_existing_version(&self) -> Result<i64> {
58 let mut version_query = self
59 .db
60 .prepare("SELECT version from semantic_index_config")?;
61 version_query
62 .query_row([], |row| Ok(row.get::<_, i64>(0)?))
63 .map_err(|err| anyhow!("version query failed: {err}"))
64 }
65
66 fn initialize_database(&self) -> Result<()> {
67 rusqlite::vtab::array::load_module(&self.db)?;
68
69 // Delete existing tables, if SEMANTIC_INDEX_VERSION is bumped
70 if self
71 .get_existing_version()
72 .map_or(false, |version| version == SEMANTIC_INDEX_VERSION as i64)
73 {
74 log::trace!("vector database schema up to date");
75 return Ok(());
76 }
77
78 log::trace!("vector database schema out of date. updating...");
79 self.db
80 .execute("DROP TABLE IF EXISTS documents", [])
81 .context("failed to drop 'documents' table")?;
82 self.db
83 .execute("DROP TABLE IF EXISTS files", [])
84 .context("failed to drop 'files' table")?;
85 self.db
86 .execute("DROP TABLE IF EXISTS worktrees", [])
87 .context("failed to drop 'worktrees' table")?;
88 self.db
89 .execute("DROP TABLE IF EXISTS semantic_index_config", [])
90 .context("failed to drop 'semantic_index_config' table")?;
91
92 // Initialize Vector Databasing Tables
93 self.db.execute(
94 "CREATE TABLE semantic_index_config (
95 version INTEGER NOT NULL
96 )",
97 [],
98 )?;
99
100 self.db.execute(
101 "INSERT INTO semantic_index_config (version) VALUES (?1)",
102 params![SEMANTIC_INDEX_VERSION],
103 )?;
104
105 self.db.execute(
106 "CREATE TABLE worktrees (
107 id INTEGER PRIMARY KEY AUTOINCREMENT,
108 absolute_path VARCHAR NOT NULL
109 );
110 CREATE UNIQUE INDEX worktrees_absolute_path ON worktrees (absolute_path);
111 ",
112 [],
113 )?;
114
115 self.db.execute(
116 "CREATE TABLE files (
117 id INTEGER PRIMARY KEY AUTOINCREMENT,
118 worktree_id INTEGER NOT NULL,
119 relative_path VARCHAR NOT NULL,
120 mtime_seconds INTEGER NOT NULL,
121 mtime_nanos INTEGER NOT NULL,
122 FOREIGN KEY(worktree_id) REFERENCES worktrees(id) ON DELETE CASCADE
123 )",
124 [],
125 )?;
126
127 self.db.execute(
128 "CREATE TABLE documents (
129 id INTEGER PRIMARY KEY AUTOINCREMENT,
130 file_id INTEGER NOT NULL,
131 start_byte INTEGER NOT NULL,
132 end_byte INTEGER NOT NULL,
133 name VARCHAR NOT NULL,
134 embedding BLOB NOT NULL,
135 FOREIGN KEY(file_id) REFERENCES files(id) ON DELETE CASCADE
136 )",
137 [],
138 )?;
139
140 log::trace!("vector database initialized with updated schema.");
141 Ok(())
142 }
143
144 pub fn delete_file(&self, worktree_id: i64, delete_path: PathBuf) -> Result<()> {
145 self.db.execute(
146 "DELETE FROM files WHERE worktree_id = ?1 AND relative_path = ?2",
147 params![worktree_id, delete_path.to_str()],
148 )?;
149 Ok(())
150 }
151
152 pub fn insert_file(
153 &self,
154 worktree_id: i64,
155 path: PathBuf,
156 mtime: SystemTime,
157 documents: Vec<Document>,
158 ) -> Result<()> {
159 // Write to files table, and return generated id.
160 self.db.execute(
161 "
162 DELETE FROM files WHERE worktree_id = ?1 AND relative_path = ?2;
163 ",
164 params![worktree_id, path.to_str()],
165 )?;
166 let mtime = Timestamp::from(mtime);
167 self.db.execute(
168 "
169 INSERT INTO files
170 (worktree_id, relative_path, mtime_seconds, mtime_nanos)
171 VALUES
172 (?1, ?2, $3, $4);
173 ",
174 params![worktree_id, path.to_str(), mtime.seconds, mtime.nanos],
175 )?;
176
177 let file_id = self.db.last_insert_rowid();
178
179 // Currently inserting at approximately 3400 documents a second
180 // I imagine we can speed this up with a bulk insert of some kind.
181 for document in documents {
182 let embedding_blob = bincode::serialize(&document.embedding)?;
183
184 self.db.execute(
185 "INSERT INTO documents (file_id, start_byte, end_byte, name, embedding) VALUES (?1, ?2, ?3, ?4, $5)",
186 params![
187 file_id,
188 document.range.start.to_string(),
189 document.range.end.to_string(),
190 document.name,
191 embedding_blob
192 ],
193 )?;
194 }
195
196 Ok(())
197 }
198
199 pub fn find_or_create_worktree(&self, worktree_root_path: &Path) -> Result<i64> {
200 // Check that the absolute path doesnt exist
201 let mut worktree_query = self
202 .db
203 .prepare("SELECT id FROM worktrees WHERE absolute_path = ?1")?;
204
205 let worktree_id = worktree_query
206 .query_row(params![worktree_root_path.to_string_lossy()], |row| {
207 Ok(row.get::<_, i64>(0)?)
208 })
209 .map_err(|err| anyhow!(err));
210
211 if worktree_id.is_ok() {
212 return worktree_id;
213 }
214
215 // If worktree_id is Err, insert new worktree
216 self.db.execute(
217 "
218 INSERT into worktrees (absolute_path) VALUES (?1)
219 ",
220 params![worktree_root_path.to_string_lossy()],
221 )?;
222 Ok(self.db.last_insert_rowid())
223 }
224
225 pub fn get_file_mtimes(&self, worktree_id: i64) -> Result<HashMap<PathBuf, SystemTime>> {
226 let mut statement = self.db.prepare(
227 "
228 SELECT relative_path, mtime_seconds, mtime_nanos
229 FROM files
230 WHERE worktree_id = ?1
231 ORDER BY relative_path",
232 )?;
233 let mut result: HashMap<PathBuf, SystemTime> = HashMap::new();
234 for row in statement.query_map(params![worktree_id], |row| {
235 Ok((
236 row.get::<_, String>(0)?.into(),
237 Timestamp {
238 seconds: row.get(1)?,
239 nanos: row.get(2)?,
240 }
241 .into(),
242 ))
243 })? {
244 let row = row?;
245 result.insert(row.0, row.1);
246 }
247 Ok(result)
248 }
249
250 pub fn top_k_search(
251 &self,
252 worktree_ids: &[i64],
253 query_embedding: &Vec<f32>,
254 limit: usize,
255 ) -> Result<Vec<(i64, PathBuf, Range<usize>)>> {
256 let mut results = Vec::<(i64, f32)>::with_capacity(limit + 1);
257 self.for_each_document(&worktree_ids, |id, embedding| {
258 let similarity = dot(&embedding, &query_embedding);
259 let ix = match results
260 .binary_search_by(|(_, s)| similarity.partial_cmp(&s).unwrap_or(Ordering::Equal))
261 {
262 Ok(ix) => ix,
263 Err(ix) => ix,
264 };
265 results.insert(ix, (id, similarity));
266 results.truncate(limit);
267 })?;
268
269 let ids = results.into_iter().map(|(id, _)| id).collect::<Vec<_>>();
270 self.get_documents_by_ids(&ids)
271 }
272
273 fn for_each_document(
274 &self,
275 worktree_ids: &[i64],
276 mut f: impl FnMut(i64, Vec<f32>),
277 ) -> Result<()> {
278 let mut query_statement = self.db.prepare(
279 "
280 SELECT
281 documents.id, documents.embedding
282 FROM
283 documents, files
284 WHERE
285 documents.file_id = files.id AND
286 files.worktree_id IN rarray(?)
287 ",
288 )?;
289
290 query_statement
291 .query_map(params![ids_to_sql(worktree_ids)], |row| {
292 Ok((row.get(0)?, row.get::<_, Embedding>(1)?))
293 })?
294 .filter_map(|row| row.ok())
295 .for_each(|(id, embedding)| f(id, embedding.0));
296 Ok(())
297 }
298
299 fn get_documents_by_ids(&self, ids: &[i64]) -> Result<Vec<(i64, PathBuf, Range<usize>)>> {
300 let mut statement = self.db.prepare(
301 "
302 SELECT
303 documents.id,
304 files.worktree_id,
305 files.relative_path,
306 documents.start_byte,
307 documents.end_byte
308 FROM
309 documents, files
310 WHERE
311 documents.file_id = files.id AND
312 documents.id in rarray(?)
313 ",
314 )?;
315
316 let result_iter = statement.query_map(params![ids_to_sql(ids)], |row| {
317 Ok((
318 row.get::<_, i64>(0)?,
319 row.get::<_, i64>(1)?,
320 row.get::<_, String>(2)?.into(),
321 row.get(3)?..row.get(4)?,
322 ))
323 })?;
324
325 let mut values_by_id = HashMap::<i64, (i64, PathBuf, Range<usize>)>::default();
326 for row in result_iter {
327 let (id, worktree_id, path, range) = row?;
328 values_by_id.insert(id, (worktree_id, path, range));
329 }
330
331 let mut results = Vec::with_capacity(ids.len());
332 for id in ids {
333 let value = values_by_id
334 .remove(id)
335 .ok_or(anyhow!("missing document id {}", id))?;
336 results.push(value);
337 }
338
339 Ok(results)
340 }
341}
342
343fn ids_to_sql(ids: &[i64]) -> Rc<Vec<rusqlite::types::Value>> {
344 Rc::new(
345 ids.iter()
346 .copied()
347 .map(|v| rusqlite::types::Value::from(v))
348 .collect::<Vec<_>>(),
349 )
350}
351
352pub(crate) fn dot(vec_a: &[f32], vec_b: &[f32]) -> f32 {
353 let len = vec_a.len();
354 assert_eq!(len, vec_b.len());
355
356 let mut result = 0.0;
357 unsafe {
358 matrixmultiply::sgemm(
359 1,
360 len,
361 1,
362 1.0,
363 vec_a.as_ptr(),
364 len as isize,
365 1,
366 vec_b.as_ptr(),
367 1,
368 len as isize,
369 0.0,
370 &mut result as *mut f32,
371 1,
372 1,
373 );
374 }
375 result
376}