1use std::collections::HashMap;
2use std::path::{Path, PathBuf};
3
4use md5::{Digest, Md5};
5use serde::{Deserialize, Serialize};
6
7#[derive(Debug, Clone, Serialize, Deserialize)]
8pub struct CodeChunk {
9 pub file_path: String,
10 pub symbol_name: String,
11 pub kind: ChunkKind,
12 pub start_line: usize,
13 pub end_line: usize,
14 pub content: String,
15 pub tokens: Vec<String>,
16 pub token_count: usize,
17}
18
19#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
20pub enum ChunkKind {
21 Function,
22 Struct,
23 Impl,
24 Module,
25 Class,
26 Method,
27 Other,
28}
29
30#[derive(Debug, Clone, Serialize, Deserialize)]
31pub struct BM25Index {
32 pub chunks: Vec<CodeChunk>,
33 pub inverted: HashMap<String, Vec<(usize, f64)>>,
34 pub avg_doc_len: f64,
35 pub doc_count: usize,
36 pub doc_freqs: HashMap<String, usize>,
37}
38
39#[derive(Debug, Clone)]
40pub struct SearchResult {
41 pub chunk_idx: usize,
42 pub score: f64,
43 pub file_path: String,
44 pub symbol_name: String,
45 pub kind: ChunkKind,
46 pub start_line: usize,
47 pub end_line: usize,
48 pub snippet: String,
49}
50
51const BM25_K1: f64 = 1.2;
52const BM25_B: f64 = 0.75;
53
54impl Default for BM25Index {
55 fn default() -> Self {
56 Self::new()
57 }
58}
59
60impl BM25Index {
61 pub fn new() -> Self {
62 Self {
63 chunks: Vec::new(),
64 inverted: HashMap::new(),
65 avg_doc_len: 0.0,
66 doc_count: 0,
67 doc_freqs: HashMap::new(),
68 }
69 }
70
71 pub fn build_from_directory(root: &Path) -> Self {
72 let mut index = Self::new();
73 let walker = ignore::WalkBuilder::new(root)
74 .hidden(true)
75 .git_ignore(true)
76 .max_depth(Some(10))
77 .build();
78
79 let mut file_count = 0usize;
80 for entry in walker.flatten() {
81 if file_count >= 2000 {
82 break;
83 }
84 let path = entry.path();
85 if !path.is_file() {
86 continue;
87 }
88 if !is_code_file(path) {
89 continue;
90 }
91 if let Ok(content) = std::fs::read_to_string(path) {
92 let rel = path
93 .strip_prefix(root)
94 .unwrap_or(path)
95 .to_string_lossy()
96 .to_string();
97 let chunks = extract_chunks(&rel, &content);
98 for chunk in chunks {
99 index.add_chunk(chunk);
100 }
101 file_count += 1;
102 }
103 }
104
105 index.finalize();
106 index
107 }
108
109 fn add_chunk(&mut self, chunk: CodeChunk) {
110 let idx = self.chunks.len();
111
112 for token in &chunk.tokens {
113 let lower = token.to_lowercase();
114 self.inverted.entry(lower).or_default().push((idx, 1.0));
115 }
116
117 self.chunks.push(chunk);
118 }
119
120 fn finalize(&mut self) {
121 self.doc_count = self.chunks.len();
122 if self.doc_count == 0 {
123 return;
124 }
125
126 let total_len: usize = self.chunks.iter().map(|c| c.token_count).sum();
127 self.avg_doc_len = total_len as f64 / self.doc_count as f64;
128
129 self.doc_freqs.clear();
130 for (term, postings) in &self.inverted {
131 let unique_docs: std::collections::HashSet<usize> =
132 postings.iter().map(|(idx, _)| *idx).collect();
133 self.doc_freqs.insert(term.clone(), unique_docs.len());
134 }
135 }
136
137 pub fn search(&self, query: &str, top_k: usize) -> Vec<SearchResult> {
138 let query_tokens = tokenize(query);
139 if query_tokens.is_empty() || self.doc_count == 0 {
140 return Vec::new();
141 }
142
143 let mut scores: HashMap<usize, f64> = HashMap::new();
144
145 for token in &query_tokens {
146 let lower = token.to_lowercase();
147 let df = *self.doc_freqs.get(&lower).unwrap_or(&0) as f64;
148 if df == 0.0 {
149 continue;
150 }
151
152 let idf = ((self.doc_count as f64 - df + 0.5) / (df + 0.5) + 1.0).ln();
153
154 if let Some(postings) = self.inverted.get(&lower) {
155 let mut doc_tfs: HashMap<usize, f64> = HashMap::new();
156 for (idx, weight) in postings {
157 *doc_tfs.entry(*idx).or_insert(0.0) += weight;
158 }
159
160 for (doc_idx, tf) in &doc_tfs {
161 let doc_len = self.chunks[*doc_idx].token_count as f64;
162 let norm_len = doc_len / self.avg_doc_len.max(1.0);
163 let bm25 = idf * (tf * (BM25_K1 + 1.0))
164 / (tf + BM25_K1 * (1.0 - BM25_B + BM25_B * norm_len));
165
166 *scores.entry(*doc_idx).or_insert(0.0) += bm25;
167 }
168 }
169 }
170
171 let mut results: Vec<SearchResult> = scores
172 .into_iter()
173 .map(|(idx, score)| {
174 let chunk = &self.chunks[idx];
175 let snippet = chunk.content.lines().take(5).collect::<Vec<_>>().join("\n");
176 SearchResult {
177 chunk_idx: idx,
178 score,
179 file_path: chunk.file_path.clone(),
180 symbol_name: chunk.symbol_name.clone(),
181 kind: chunk.kind.clone(),
182 start_line: chunk.start_line,
183 end_line: chunk.end_line,
184 snippet,
185 }
186 })
187 .collect();
188
189 results.sort_by(|a, b| {
190 b.score
191 .partial_cmp(&a.score)
192 .unwrap_or(std::cmp::Ordering::Equal)
193 });
194 results.truncate(top_k);
195 results
196 }
197
198 pub fn save(&self, root: &Path) -> std::io::Result<()> {
199 let dir = index_dir(root);
200 std::fs::create_dir_all(&dir)?;
201 let data = serde_json::to_string(self).map_err(std::io::Error::other)?;
202 std::fs::write(dir.join("bm25_index.json"), data)?;
203 Ok(())
204 }
205
206 pub fn load(root: &Path) -> Option<Self> {
207 let path = index_dir(root).join("bm25_index.json");
208 let data = std::fs::read_to_string(path).ok()?;
209 serde_json::from_str(&data).ok()
210 }
211
212 pub fn load_or_build(root: &Path) -> Self {
213 Self::load(root).unwrap_or_else(|| {
214 let built = Self::build_from_directory(root);
215 let _ = built.save(root);
216 built
217 })
218 }
219
220 pub fn index_file_path(root: &Path) -> PathBuf {
221 index_dir(root).join("bm25_index.json")
222 }
223}
224
225fn index_dir(root: &Path) -> PathBuf {
226 let mut hasher = Md5::new();
227 hasher.update(root.to_string_lossy().as_bytes());
228 let hash = format!("{:x}", hasher.finalize());
229 crate::core::data_dir::lean_ctx_data_dir()
230 .unwrap_or_else(|_| PathBuf::from("."))
231 .join("vectors")
232 .join(hash)
233}
234
235pub(crate) fn is_code_file(path: &Path) -> bool {
236 let ext = path.extension().and_then(|e| e.to_str()).unwrap_or("");
237 matches!(
238 ext,
239 "rs" | "ts"
240 | "tsx"
241 | "js"
242 | "jsx"
243 | "py"
244 | "go"
245 | "java"
246 | "c"
247 | "cpp"
248 | "h"
249 | "hpp"
250 | "rb"
251 | "cs"
252 | "kt"
253 | "swift"
254 | "php"
255 | "scala"
256 | "ex"
257 | "exs"
258 | "zig"
259 | "lua"
260 | "dart"
261 | "vue"
262 | "svelte"
263 )
264}
265
266fn tokenize(text: &str) -> Vec<String> {
267 let mut tokens = Vec::new();
268 let mut current = String::new();
269
270 for ch in text.chars() {
271 if ch.is_alphanumeric() || ch == '_' {
272 current.push(ch);
273 } else {
274 if current.len() >= 2 {
275 tokens.push(current.clone());
276 }
277 current.clear();
278 }
279 }
280 if current.len() >= 2 {
281 tokens.push(current);
282 }
283
284 split_camel_case_tokens(&tokens)
285}
286
287fn split_camel_case_tokens(tokens: &[String]) -> Vec<String> {
288 let mut result = Vec::new();
289 for token in tokens {
290 result.push(token.clone());
291 let mut start = 0;
292 let chars: Vec<char> = token.chars().collect();
293 for i in 1..chars.len() {
294 if chars[i].is_uppercase() && (i + 1 >= chars.len() || !chars[i + 1].is_uppercase()) {
295 let part: String = chars[start..i].iter().collect();
296 if part.len() >= 2 {
297 result.push(part);
298 }
299 start = i;
300 }
301 }
302 if start > 0 {
303 let part: String = chars[start..].iter().collect();
304 if part.len() >= 2 {
305 result.push(part);
306 }
307 }
308 }
309 result
310}
311
312fn extract_chunks(file_path: &str, content: &str) -> Vec<CodeChunk> {
313 let lines: Vec<&str> = content.lines().collect();
314 if lines.is_empty() {
315 return Vec::new();
316 }
317
318 let mut chunks = Vec::new();
319 let mut i = 0;
320
321 while i < lines.len() {
322 let trimmed = lines[i].trim();
323
324 if let Some((name, kind)) = detect_symbol(trimmed) {
325 let start = i;
326 let end = find_block_end(&lines, i);
327 let block: String = lines[start..=end.min(lines.len() - 1)].to_vec().join("\n");
328 let tokens = tokenize(&block);
329 let token_count = tokens.len();
330
331 chunks.push(CodeChunk {
332 file_path: file_path.to_string(),
333 symbol_name: name,
334 kind,
335 start_line: start + 1,
336 end_line: end + 1,
337 content: block,
338 tokens,
339 token_count,
340 });
341
342 i = end + 1;
343 } else {
344 i += 1;
345 }
346 }
347
348 if chunks.is_empty() && !content.is_empty() {
349 let tokens = tokenize(content);
350 let token_count = tokens.len();
351 let snippet = lines
352 .iter()
353 .take(50)
354 .copied()
355 .collect::<Vec<_>>()
356 .join("\n");
357 chunks.push(CodeChunk {
358 file_path: file_path.to_string(),
359 symbol_name: file_path.to_string(),
360 kind: ChunkKind::Module,
361 start_line: 1,
362 end_line: lines.len(),
363 content: snippet,
364 tokens,
365 token_count,
366 });
367 }
368
369 chunks
370}
371
372fn detect_symbol(line: &str) -> Option<(String, ChunkKind)> {
373 let trimmed = line.trim();
374
375 let patterns: &[(&str, ChunkKind)] = &[
376 ("pub async fn ", ChunkKind::Function),
377 ("async fn ", ChunkKind::Function),
378 ("pub fn ", ChunkKind::Function),
379 ("fn ", ChunkKind::Function),
380 ("pub struct ", ChunkKind::Struct),
381 ("struct ", ChunkKind::Struct),
382 ("pub enum ", ChunkKind::Struct),
383 ("enum ", ChunkKind::Struct),
384 ("impl ", ChunkKind::Impl),
385 ("pub trait ", ChunkKind::Struct),
386 ("trait ", ChunkKind::Struct),
387 ("export function ", ChunkKind::Function),
388 ("export async function ", ChunkKind::Function),
389 ("export default function ", ChunkKind::Function),
390 ("function ", ChunkKind::Function),
391 ("async function ", ChunkKind::Function),
392 ("export class ", ChunkKind::Class),
393 ("class ", ChunkKind::Class),
394 ("export interface ", ChunkKind::Struct),
395 ("interface ", ChunkKind::Struct),
396 ("def ", ChunkKind::Function),
397 ("async def ", ChunkKind::Function),
398 ("class ", ChunkKind::Class),
399 ("func ", ChunkKind::Function),
400 ];
401
402 for (prefix, kind) in patterns {
403 if let Some(rest) = trimmed.strip_prefix(prefix) {
404 let name: String = rest
405 .chars()
406 .take_while(|c| c.is_alphanumeric() || *c == '_' || *c == '<')
407 .take_while(|c| *c != '<')
408 .collect();
409 if !name.is_empty() {
410 return Some((name, kind.clone()));
411 }
412 }
413 }
414
415 None
416}
417
418fn find_block_end(lines: &[&str], start: usize) -> usize {
419 let mut depth = 0i32;
420 let mut found_open = false;
421
422 for (i, line) in lines.iter().enumerate().skip(start) {
423 for ch in line.chars() {
424 match ch {
425 '{' | '(' if !found_open || depth > 0 => {
426 depth += 1;
427 found_open = true;
428 }
429 '}' | ')' if depth > 0 => {
430 depth -= 1;
431 if depth == 0 && found_open {
432 return i;
433 }
434 }
435 _ => {}
436 }
437 }
438
439 if found_open && depth <= 0 && i > start {
440 return i;
441 }
442
443 if !found_open && i > start + 2 {
444 let trimmed = lines[i].trim();
445 if trimmed.is_empty()
446 || (!trimmed.starts_with(' ') && !trimmed.starts_with('\t') && i > start)
447 {
448 return i.saturating_sub(1);
449 }
450 }
451 }
452
453 (start + 50).min(lines.len().saturating_sub(1))
454}
455
456pub fn format_search_results(results: &[SearchResult], compact: bool) -> String {
457 if results.is_empty() {
458 return "No results found.".to_string();
459 }
460
461 let mut out = String::new();
462 for (i, r) in results.iter().enumerate() {
463 if compact {
464 out.push_str(&format!(
465 "{}. {:.2} {}:{}-{} {:?} {}\n",
466 i + 1,
467 r.score,
468 r.file_path,
469 r.start_line,
470 r.end_line,
471 r.kind,
472 r.symbol_name,
473 ));
474 } else {
475 out.push_str(&format!(
476 "\n--- Result {} (score: {:.2}) ---\n{} :: {} [{:?}] (L{}-{})\n{}\n",
477 i + 1,
478 r.score,
479 r.file_path,
480 r.symbol_name,
481 r.kind,
482 r.start_line,
483 r.end_line,
484 r.snippet,
485 ));
486 }
487 }
488 out
489}
490
491#[cfg(test)]
492mod tests {
493 use super::*;
494
495 #[test]
496 fn tokenize_splits_code() {
497 let tokens = tokenize("fn calculate_total(items: Vec<Item>) -> f64");
498 assert!(tokens.contains(&"calculate_total".to_string()));
499 assert!(tokens.contains(&"items".to_string()));
500 assert!(tokens.contains(&"Vec".to_string()));
501 }
502
503 #[test]
504 fn camel_case_splitting() {
505 let tokens = split_camel_case_tokens(&["calculateTotal".to_string()]);
506 assert!(tokens.contains(&"calculateTotal".to_string()));
507 assert!(tokens.contains(&"calculate".to_string()));
508 assert!(tokens.contains(&"Total".to_string()));
509 }
510
511 #[test]
512 fn detect_rust_function() {
513 let (name, kind) =
514 detect_symbol("pub fn process_request(req: Request) -> Response {").unwrap();
515 assert_eq!(name, "process_request");
516 assert_eq!(kind, ChunkKind::Function);
517 }
518
519 #[test]
520 fn bm25_search_finds_relevant() {
521 let mut index = BM25Index::new();
522 index.add_chunk(CodeChunk {
523 file_path: "auth.rs".into(),
524 symbol_name: "validate_token".into(),
525 kind: ChunkKind::Function,
526 start_line: 1,
527 end_line: 10,
528 content: "fn validate_token(token: &str) -> bool { check_jwt_expiry(token) }".into(),
529 tokens: tokenize("fn validate_token token str bool check_jwt_expiry token"),
530 token_count: 8,
531 });
532 index.add_chunk(CodeChunk {
533 file_path: "db.rs".into(),
534 symbol_name: "connect_database".into(),
535 kind: ChunkKind::Function,
536 start_line: 1,
537 end_line: 5,
538 content: "fn connect_database(url: &str) -> Pool { create_pool(url) }".into(),
539 tokens: tokenize("fn connect_database url str Pool create_pool url"),
540 token_count: 7,
541 });
542 index.finalize();
543
544 let results = index.search("jwt token validation", 5);
545 assert!(!results.is_empty());
546 assert_eq!(results[0].symbol_name, "validate_token");
547 }
548}