Skip to main content

lean_ctx/core/
vector_index.rs

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
213fn index_dir(root: &Path) -> PathBuf {
214    let mut hasher = Md5::new();
215    hasher.update(root.to_string_lossy().as_bytes());
216    let hash = format!("{:x}", hasher.finalize());
217    dirs::home_dir()
218        .unwrap_or_else(|| PathBuf::from("."))
219        .join(".lean-ctx")
220        .join("vectors")
221        .join(hash)
222}
223
224fn is_code_file(path: &Path) -> bool {
225    let ext = path.extension().and_then(|e| e.to_str()).unwrap_or("");
226    matches!(
227        ext,
228        "rs" | "ts"
229            | "tsx"
230            | "js"
231            | "jsx"
232            | "py"
233            | "go"
234            | "java"
235            | "c"
236            | "cpp"
237            | "h"
238            | "hpp"
239            | "rb"
240            | "cs"
241            | "kt"
242            | "swift"
243            | "php"
244            | "scala"
245            | "ex"
246            | "exs"
247            | "zig"
248            | "lua"
249            | "dart"
250            | "vue"
251            | "svelte"
252    )
253}
254
255fn tokenize(text: &str) -> Vec<String> {
256    let mut tokens = Vec::new();
257    let mut current = String::new();
258
259    for ch in text.chars() {
260        if ch.is_alphanumeric() || ch == '_' {
261            current.push(ch);
262        } else {
263            if current.len() >= 2 {
264                tokens.push(current.clone());
265            }
266            current.clear();
267        }
268    }
269    if current.len() >= 2 {
270        tokens.push(current);
271    }
272
273    split_camel_case_tokens(&tokens)
274}
275
276fn split_camel_case_tokens(tokens: &[String]) -> Vec<String> {
277    let mut result = Vec::new();
278    for token in tokens {
279        result.push(token.clone());
280        let mut start = 0;
281        let chars: Vec<char> = token.chars().collect();
282        for i in 1..chars.len() {
283            if chars[i].is_uppercase() && (i + 1 >= chars.len() || !chars[i + 1].is_uppercase()) {
284                let part: String = chars[start..i].iter().collect();
285                if part.len() >= 2 {
286                    result.push(part);
287                }
288                start = i;
289            }
290        }
291        if start > 0 {
292            let part: String = chars[start..].iter().collect();
293            if part.len() >= 2 {
294                result.push(part);
295            }
296        }
297    }
298    result
299}
300
301fn extract_chunks(file_path: &str, content: &str) -> Vec<CodeChunk> {
302    let lines: Vec<&str> = content.lines().collect();
303    if lines.is_empty() {
304        return Vec::new();
305    }
306
307    let mut chunks = Vec::new();
308    let mut i = 0;
309
310    while i < lines.len() {
311        let trimmed = lines[i].trim();
312
313        if let Some((name, kind)) = detect_symbol(trimmed) {
314            let start = i;
315            let end = find_block_end(&lines, i);
316            let block: String = lines[start..=end.min(lines.len() - 1)].to_vec().join("\n");
317            let tokens = tokenize(&block);
318            let token_count = tokens.len();
319
320            chunks.push(CodeChunk {
321                file_path: file_path.to_string(),
322                symbol_name: name,
323                kind,
324                start_line: start + 1,
325                end_line: end + 1,
326                content: block,
327                tokens,
328                token_count,
329            });
330
331            i = end + 1;
332        } else {
333            i += 1;
334        }
335    }
336
337    if chunks.is_empty() && !content.is_empty() {
338        let tokens = tokenize(content);
339        let token_count = tokens.len();
340        let snippet = lines
341            .iter()
342            .take(50)
343            .copied()
344            .collect::<Vec<_>>()
345            .join("\n");
346        chunks.push(CodeChunk {
347            file_path: file_path.to_string(),
348            symbol_name: file_path.to_string(),
349            kind: ChunkKind::Module,
350            start_line: 1,
351            end_line: lines.len(),
352            content: snippet,
353            tokens,
354            token_count,
355        });
356    }
357
358    chunks
359}
360
361fn detect_symbol(line: &str) -> Option<(String, ChunkKind)> {
362    let trimmed = line.trim();
363
364    let patterns: &[(&str, ChunkKind)] = &[
365        ("pub async fn ", ChunkKind::Function),
366        ("async fn ", ChunkKind::Function),
367        ("pub fn ", ChunkKind::Function),
368        ("fn ", ChunkKind::Function),
369        ("pub struct ", ChunkKind::Struct),
370        ("struct ", ChunkKind::Struct),
371        ("pub enum ", ChunkKind::Struct),
372        ("enum ", ChunkKind::Struct),
373        ("impl ", ChunkKind::Impl),
374        ("pub trait ", ChunkKind::Struct),
375        ("trait ", ChunkKind::Struct),
376        ("export function ", ChunkKind::Function),
377        ("export async function ", ChunkKind::Function),
378        ("export default function ", ChunkKind::Function),
379        ("function ", ChunkKind::Function),
380        ("async function ", ChunkKind::Function),
381        ("export class ", ChunkKind::Class),
382        ("class ", ChunkKind::Class),
383        ("export interface ", ChunkKind::Struct),
384        ("interface ", ChunkKind::Struct),
385        ("def ", ChunkKind::Function),
386        ("async def ", ChunkKind::Function),
387        ("class ", ChunkKind::Class),
388        ("func ", ChunkKind::Function),
389    ];
390
391    for (prefix, kind) in patterns {
392        if let Some(rest) = trimmed.strip_prefix(prefix) {
393            let name: String = rest
394                .chars()
395                .take_while(|c| c.is_alphanumeric() || *c == '_' || *c == '<')
396                .take_while(|c| *c != '<')
397                .collect();
398            if !name.is_empty() {
399                return Some((name, kind.clone()));
400            }
401        }
402    }
403
404    None
405}
406
407fn find_block_end(lines: &[&str], start: usize) -> usize {
408    let mut depth = 0i32;
409    let mut found_open = false;
410
411    for (i, line) in lines.iter().enumerate().skip(start) {
412        for ch in line.chars() {
413            match ch {
414                '{' | '(' if !found_open || depth > 0 => {
415                    depth += 1;
416                    found_open = true;
417                }
418                '}' | ')' if depth > 0 => {
419                    depth -= 1;
420                    if depth == 0 && found_open {
421                        return i;
422                    }
423                }
424                _ => {}
425            }
426        }
427
428        if found_open && depth <= 0 && i > start {
429            return i;
430        }
431
432        if !found_open && i > start + 2 {
433            let trimmed = lines[i].trim();
434            if trimmed.is_empty()
435                || (!trimmed.starts_with(' ') && !trimmed.starts_with('\t') && i > start)
436            {
437                return i.saturating_sub(1);
438            }
439        }
440    }
441
442    (start + 50).min(lines.len().saturating_sub(1))
443}
444
445pub fn format_search_results(results: &[SearchResult], compact: bool) -> String {
446    if results.is_empty() {
447        return "No results found.".to_string();
448    }
449
450    let mut out = String::new();
451    for (i, r) in results.iter().enumerate() {
452        if compact {
453            out.push_str(&format!(
454                "{}. {:.2} {}:{}-{} {:?} {}\n",
455                i + 1,
456                r.score,
457                r.file_path,
458                r.start_line,
459                r.end_line,
460                r.kind,
461                r.symbol_name,
462            ));
463        } else {
464            out.push_str(&format!(
465                "\n--- Result {} (score: {:.2}) ---\n{} :: {} [{:?}] (L{}-{})\n{}\n",
466                i + 1,
467                r.score,
468                r.file_path,
469                r.symbol_name,
470                r.kind,
471                r.start_line,
472                r.end_line,
473                r.snippet,
474            ));
475        }
476    }
477    out
478}
479
480#[cfg(test)]
481mod tests {
482    use super::*;
483
484    #[test]
485    fn tokenize_splits_code() {
486        let tokens = tokenize("fn calculate_total(items: Vec<Item>) -> f64");
487        assert!(tokens.contains(&"calculate_total".to_string()));
488        assert!(tokens.contains(&"items".to_string()));
489        assert!(tokens.contains(&"Vec".to_string()));
490    }
491
492    #[test]
493    fn camel_case_splitting() {
494        let tokens = split_camel_case_tokens(&["calculateTotal".to_string()]);
495        assert!(tokens.contains(&"calculateTotal".to_string()));
496        assert!(tokens.contains(&"calculate".to_string()));
497        assert!(tokens.contains(&"Total".to_string()));
498    }
499
500    #[test]
501    fn detect_rust_function() {
502        let (name, kind) =
503            detect_symbol("pub fn process_request(req: Request) -> Response {").unwrap();
504        assert_eq!(name, "process_request");
505        assert_eq!(kind, ChunkKind::Function);
506    }
507
508    #[test]
509    fn bm25_search_finds_relevant() {
510        let mut index = BM25Index::new();
511        index.add_chunk(CodeChunk {
512            file_path: "auth.rs".into(),
513            symbol_name: "validate_token".into(),
514            kind: ChunkKind::Function,
515            start_line: 1,
516            end_line: 10,
517            content: "fn validate_token(token: &str) -> bool { check_jwt_expiry(token) }".into(),
518            tokens: tokenize("fn validate_token token str bool check_jwt_expiry token"),
519            token_count: 8,
520        });
521        index.add_chunk(CodeChunk {
522            file_path: "db.rs".into(),
523            symbol_name: "connect_database".into(),
524            kind: ChunkKind::Function,
525            start_line: 1,
526            end_line: 5,
527            content: "fn connect_database(url: &str) -> Pool { create_pool(url) }".into(),
528            tokens: tokenize("fn connect_database url str Pool create_pool url"),
529            token_count: 7,
530        });
531        index.finalize();
532
533        let results = index.search("jwt token validation", 5);
534        assert!(!results.is_empty());
535        assert_eq!(results[0].symbol_name, "validate_token");
536    }
537}