1use crate::{parsing::Document, SEMANTIC_INDEX_VERSION};
2use anyhow::{anyhow, Context, Result};
3use globset::GlobMatcher;
4use project::Fs;
5use rpc::proto::Timestamp;
6use rusqlite::{
7 params,
8 types::{FromSql, FromSqlResult, ValueRef},
9};
10use std::{
11 cmp::Ordering,
12 collections::HashMap,
13 ops::Range,
14 path::{Path, PathBuf},
15 rc::Rc,
16 sync::Arc,
17 time::SystemTime,
18};
19
20#[derive(Debug)]
21pub struct FileRecord {
22 pub id: usize,
23 pub relative_path: String,
24 pub mtime: Timestamp,
25}
26
27#[derive(Debug)]
28struct Embedding(pub Vec<f32>);
29
30impl FromSql for Embedding {
31 fn column_result(value: ValueRef) -> FromSqlResult<Self> {
32 let bytes = value.as_blob()?;
33 let embedding: Result<Vec<f32>, Box<bincode::ErrorKind>> = bincode::deserialize(bytes);
34 if embedding.is_err() {
35 return Err(rusqlite::types::FromSqlError::Other(embedding.unwrap_err()));
36 }
37 return Ok(Embedding(embedding.unwrap()));
38 }
39}
40
41pub struct VectorDatabase {
42 db: rusqlite::Connection,
43}
44
45impl VectorDatabase {
46 pub async fn new(fs: Arc<dyn Fs>, path: Arc<PathBuf>) -> Result<Self> {
47 if let Some(db_directory) = path.parent() {
48 fs.create_dir(db_directory).await?;
49 }
50
51 let this = Self {
52 db: rusqlite::Connection::open(path.as_path())?,
53 };
54 this.initialize_database()?;
55 Ok(this)
56 }
57
58 fn get_existing_version(&self) -> Result<i64> {
59 let mut version_query = self
60 .db
61 .prepare("SELECT version from semantic_index_config")?;
62 version_query
63 .query_row([], |row| Ok(row.get::<_, i64>(0)?))
64 .map_err(|err| anyhow!("version query failed: {err}"))
65 }
66
67 fn initialize_database(&self) -> Result<()> {
68 rusqlite::vtab::array::load_module(&self.db)?;
69
70 // Delete existing tables, if SEMANTIC_INDEX_VERSION is bumped
71 if self
72 .get_existing_version()
73 .map_or(false, |version| version == SEMANTIC_INDEX_VERSION as i64)
74 {
75 log::trace!("vector database schema up to date");
76 return Ok(());
77 }
78
79 log::trace!("vector database schema out of date. updating...");
80 self.db
81 .execute("DROP TABLE IF EXISTS documents", [])
82 .context("failed to drop 'documents' table")?;
83 self.db
84 .execute("DROP TABLE IF EXISTS files", [])
85 .context("failed to drop 'files' table")?;
86 self.db
87 .execute("DROP TABLE IF EXISTS worktrees", [])
88 .context("failed to drop 'worktrees' table")?;
89 self.db
90 .execute("DROP TABLE IF EXISTS semantic_index_config", [])
91 .context("failed to drop 'semantic_index_config' table")?;
92
93 // Initialize Vector Databasing Tables
94 self.db.execute(
95 "CREATE TABLE semantic_index_config (
96 version INTEGER NOT NULL
97 )",
98 [],
99 )?;
100
101 self.db.execute(
102 "INSERT INTO semantic_index_config (version) VALUES (?1)",
103 params![SEMANTIC_INDEX_VERSION],
104 )?;
105
106 self.db.execute(
107 "CREATE TABLE worktrees (
108 id INTEGER PRIMARY KEY AUTOINCREMENT,
109 absolute_path VARCHAR NOT NULL
110 );
111 CREATE UNIQUE INDEX worktrees_absolute_path ON worktrees (absolute_path);
112 ",
113 [],
114 )?;
115
116 self.db.execute(
117 "CREATE TABLE files (
118 id INTEGER PRIMARY KEY AUTOINCREMENT,
119 worktree_id INTEGER NOT NULL,
120 relative_path VARCHAR NOT NULL,
121 mtime_seconds INTEGER NOT NULL,
122 mtime_nanos INTEGER NOT NULL,
123 FOREIGN KEY(worktree_id) REFERENCES worktrees(id) ON DELETE CASCADE
124 )",
125 [],
126 )?;
127
128 self.db.execute(
129 "CREATE TABLE documents (
130 id INTEGER PRIMARY KEY AUTOINCREMENT,
131 file_id INTEGER NOT NULL,
132 start_byte INTEGER NOT NULL,
133 end_byte INTEGER NOT NULL,
134 name VARCHAR NOT NULL,
135 embedding BLOB NOT NULL,
136 FOREIGN KEY(file_id) REFERENCES files(id) ON DELETE CASCADE
137 )",
138 [],
139 )?;
140
141 log::trace!("vector database initialized with updated schema.");
142 Ok(())
143 }
144
145 pub fn delete_file(&self, worktree_id: i64, delete_path: PathBuf) -> Result<()> {
146 self.db.execute(
147 "DELETE FROM files WHERE worktree_id = ?1 AND relative_path = ?2",
148 params![worktree_id, delete_path.to_str()],
149 )?;
150 Ok(())
151 }
152
153 pub fn insert_file(
154 &self,
155 worktree_id: i64,
156 path: PathBuf,
157 mtime: SystemTime,
158 documents: Vec<Document>,
159 ) -> Result<()> {
160 // Write to files table, and return generated id.
161 self.db.execute(
162 "
163 DELETE FROM files WHERE worktree_id = ?1 AND relative_path = ?2;
164 ",
165 params![worktree_id, path.to_str()],
166 )?;
167 let mtime = Timestamp::from(mtime);
168 self.db.execute(
169 "
170 INSERT INTO files
171 (worktree_id, relative_path, mtime_seconds, mtime_nanos)
172 VALUES
173 (?1, ?2, $3, $4);
174 ",
175 params![worktree_id, path.to_str(), mtime.seconds, mtime.nanos],
176 )?;
177
178 let file_id = self.db.last_insert_rowid();
179
180 // Currently inserting at approximately 3400 documents a second
181 // I imagine we can speed this up with a bulk insert of some kind.
182 for document in documents {
183 let embedding_blob = bincode::serialize(&document.embedding)?;
184
185 self.db.execute(
186 "INSERT INTO documents (file_id, start_byte, end_byte, name, embedding) VALUES (?1, ?2, ?3, ?4, $5)",
187 params![
188 file_id,
189 document.range.start.to_string(),
190 document.range.end.to_string(),
191 document.name,
192 embedding_blob
193 ],
194 )?;
195 }
196
197 Ok(())
198 }
199
200 pub fn worktree_previously_indexed(&self, worktree_root_path: &Path) -> Result<bool> {
201 let mut worktree_query = self
202 .db
203 .prepare("SELECT id FROM worktrees WHERE absolute_path = ?1")?;
204 let worktree_id = worktree_query
205 .query_row(params![worktree_root_path.to_string_lossy()], |row| {
206 Ok(row.get::<_, i64>(0)?)
207 })
208 .map_err(|err| anyhow!(err));
209
210 if worktree_id.is_ok() {
211 return Ok(true);
212 } else {
213 return Ok(false);
214 }
215 }
216
217 pub fn find_or_create_worktree(&self, worktree_root_path: &Path) -> Result<i64> {
218 // Check that the absolute path doesnt exist
219 let mut worktree_query = self
220 .db
221 .prepare("SELECT id FROM worktrees WHERE absolute_path = ?1")?;
222
223 let worktree_id = worktree_query
224 .query_row(params![worktree_root_path.to_string_lossy()], |row| {
225 Ok(row.get::<_, i64>(0)?)
226 })
227 .map_err(|err| anyhow!(err));
228
229 if worktree_id.is_ok() {
230 return worktree_id;
231 }
232
233 // If worktree_id is Err, insert new worktree
234 self.db.execute(
235 "
236 INSERT into worktrees (absolute_path) VALUES (?1)
237 ",
238 params![worktree_root_path.to_string_lossy()],
239 )?;
240 Ok(self.db.last_insert_rowid())
241 }
242
243 pub fn get_file_mtimes(&self, worktree_id: i64) -> Result<HashMap<PathBuf, SystemTime>> {
244 let mut statement = self.db.prepare(
245 "
246 SELECT relative_path, mtime_seconds, mtime_nanos
247 FROM files
248 WHERE worktree_id = ?1
249 ORDER BY relative_path",
250 )?;
251 let mut result: HashMap<PathBuf, SystemTime> = HashMap::new();
252 for row in statement.query_map(params![worktree_id], |row| {
253 Ok((
254 row.get::<_, String>(0)?.into(),
255 Timestamp {
256 seconds: row.get(1)?,
257 nanos: row.get(2)?,
258 }
259 .into(),
260 ))
261 })? {
262 let row = row?;
263 result.insert(row.0, row.1);
264 }
265 Ok(result)
266 }
267
268 pub fn top_k_search(
269 &self,
270 query_embedding: &Vec<f32>,
271 limit: usize,
272 file_ids: &[i64],
273 ) -> Result<Vec<(i64, f32)>> {
274 let mut results = Vec::<(i64, f32)>::with_capacity(limit + 1);
275 self.for_each_document(file_ids, |id, embedding| {
276 let similarity = dot(&embedding, &query_embedding);
277 let ix = match results
278 .binary_search_by(|(_, s)| similarity.partial_cmp(&s).unwrap_or(Ordering::Equal))
279 {
280 Ok(ix) => ix,
281 Err(ix) => ix,
282 };
283 results.insert(ix, (id, similarity));
284 results.truncate(limit);
285 })?;
286
287 Ok(results)
288 }
289
290 pub fn retrieve_included_file_ids(
291 &self,
292 worktree_ids: &[i64],
293 include_globs: Vec<GlobMatcher>,
294 exclude_globs: Vec<GlobMatcher>,
295 ) -> Result<Vec<i64>> {
296 let mut file_query = self.db.prepare(
297 "
298 SELECT
299 id, relative_path
300 FROM
301 files
302 WHERE
303 worktree_id IN rarray(?)
304 ",
305 )?;
306
307 let mut file_ids = Vec::<i64>::new();
308 let mut rows = file_query.query([ids_to_sql(worktree_ids)])?;
309
310 while let Some(row) = rows.next()? {
311 let file_id = row.get(0)?;
312 let relative_path = row.get_ref(1)?.as_str()?;
313 let included = include_globs.is_empty()
314 || include_globs
315 .iter()
316 .any(|glob| glob.is_match(relative_path));
317 let excluded = exclude_globs
318 .iter()
319 .any(|glob| glob.is_match(relative_path));
320 if included && !excluded {
321 file_ids.push(file_id);
322 }
323 }
324
325 Ok(file_ids)
326 }
327
328 fn for_each_document(&self, file_ids: &[i64], mut f: impl FnMut(i64, Vec<f32>)) -> Result<()> {
329 let mut query_statement = self.db.prepare(
330 "
331 SELECT
332 id, embedding
333 FROM
334 documents
335 WHERE
336 file_id IN rarray(?)
337 ",
338 )?;
339
340 query_statement
341 .query_map(params![ids_to_sql(&file_ids)], |row| {
342 Ok((row.get(0)?, row.get::<_, Embedding>(1)?))
343 })?
344 .filter_map(|row| row.ok())
345 .for_each(|(id, embedding)| f(id, embedding.0));
346 Ok(())
347 }
348
349 pub fn get_documents_by_ids(&self, ids: &[i64]) -> Result<Vec<(i64, PathBuf, Range<usize>)>> {
350 let mut statement = self.db.prepare(
351 "
352 SELECT
353 documents.id,
354 files.worktree_id,
355 files.relative_path,
356 documents.start_byte,
357 documents.end_byte
358 FROM
359 documents, files
360 WHERE
361 documents.file_id = files.id AND
362 documents.id in rarray(?)
363 ",
364 )?;
365
366 let result_iter = statement.query_map(params![ids_to_sql(ids)], |row| {
367 Ok((
368 row.get::<_, i64>(0)?,
369 row.get::<_, i64>(1)?,
370 row.get::<_, String>(2)?.into(),
371 row.get(3)?..row.get(4)?,
372 ))
373 })?;
374
375 let mut values_by_id = HashMap::<i64, (i64, PathBuf, Range<usize>)>::default();
376 for row in result_iter {
377 let (id, worktree_id, path, range) = row?;
378 values_by_id.insert(id, (worktree_id, path, range));
379 }
380
381 let mut results = Vec::with_capacity(ids.len());
382 for id in ids {
383 let value = values_by_id
384 .remove(id)
385 .ok_or(anyhow!("missing document id {}", id))?;
386 results.push(value);
387 }
388
389 Ok(results)
390 }
391}
392
393fn ids_to_sql(ids: &[i64]) -> Rc<Vec<rusqlite::types::Value>> {
394 Rc::new(
395 ids.iter()
396 .copied()
397 .map(|v| rusqlite::types::Value::from(v))
398 .collect::<Vec<_>>(),
399 )
400}
401
402pub(crate) fn dot(vec_a: &[f32], vec_b: &[f32]) -> f32 {
403 let len = vec_a.len();
404 assert_eq!(len, vec_b.len());
405
406 let mut result = 0.0;
407 unsafe {
408 matrixmultiply::sgemm(
409 1,
410 len,
411 1,
412 1.0,
413 vec_a.as_ptr(),
414 len as isize,
415 1,
416 vec_b.as_ptr(),
417 1,
418 len as isize,
419 0.0,
420 &mut result as *mut f32,
421 1,
422 1,
423 );
424 }
425 result
426}