rlm-cli 1.2.4

Recursive Language Model (RLM) REPL for Claude Code - handles long-context tasks via chunking and recursive sub-LLM calls
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
# RLM-RS Architecture

Internal architecture documentation for `rlm-rs`.

## Overview

RLM-RS implements the Recursive Language Model (RLM) pattern from [arXiv:2512.24601](https://arxiv.org/abs/2512.24601), enabling LLMs to process documents up to 100x larger than their context windows.

## Architecture Diagram

```
┌─────────────────────────────────────────────────────────────────┐
│                        Claude Code                               │
│  ┌─────────────────┐    ┌─────────────────┐                     │
│  │   Root LLM      │───▶│   Sub-LLM       │                     │
│  │ (Opus/Sonnet)   │    │   (Haiku)       │                     │
│  └────────┬────────┘    └────────┬────────┘                     │
│           │                      │                               │
│           ▼                      ▼                               │
│  ┌─────────────────────────────────────────┐                    │
│  │              Bash Tool                   │                    │
│  └─────────────────┬───────────────────────┘                    │
└────────────────────┼────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────┐
│                        rlm-rs CLI                                │
│  ┌─────────────────────────────────────────────────────────────┐│
│  │                      CLI Layer                               ││
│  │  parser.rs │ commands.rs │ output.rs                        ││
│  └─────────────────────────┬───────────────────────────────────┘│
│                            │                                     │
│  ┌─────────────────────────┴───────────────────────────────────┐│
│  │                    Core Domain                               ││
│  │  Buffer │ Chunk │ Context │ Variable                        ││
│  └─────────────────────────┬───────────────────────────────────┘│
│                            │                                     │
│  ┌────────────┬────────────┴────────────┬────────────┬────────────┐│
│  │  Chunking  │       Storage           │ Embedding  │    I/O     ││
│  │  ─────────  │       ───────           │ ─────────  │    ───     ││
│  │  Fixed     │       SQLite            │  BGE-M3   │   Reader   ││
│  │  Semantic  │       FTS5 (BM25)       │ fastembed │   (mmap)   ││
│  │  Code      │       Hybrid Search     │  (1024d)  │   Unicode  ││
│  │  Parallel  │       HNSW (optional)   │           │            ││
│  └────────────┴─────────────────────────┴────────────┴────────────┘│
└─────────────────────────────────────────────────────────────────────┘
```

## Module Structure

```
src/
├── lib.rs           # Library entry point and public API
├── main.rs          # Binary entry point
├── error.rs         # Error types (thiserror)
│
├── core/            # Core domain types
│   ├── mod.rs
│   ├── buffer.rs    # Buffer: loaded file content
│   ├── chunk.rs     # Chunk: content segment with metadata
│   └── context.rs   # Context: variables and state
│
├── chunking/        # Chunking strategies
│   ├── mod.rs       # Strategy factory and constants
│   ├── traits.rs    # Chunker trait definition
│   ├── fixed.rs     # Fixed-size chunking
│   ├── semantic.rs  # Sentence/paragraph-aware chunking
│   ├── code.rs      # Language-aware code chunking
│   └── parallel.rs  # Multi-threaded chunking
│
├── embedding/       # Embedding generation
│   ├── mod.rs       # Embedding trait and constants
│   ├── fastembed_impl.rs  # BGE-M3 via fastembed-rs
│   └── fallback.rs  # Fallback when fastembed unavailable
│
├── storage/         # Persistence layer
│   ├── mod.rs
│   ├── traits.rs    # Storage trait definition
│   ├── sqlite.rs    # SQLite implementation
│   └── search.rs    # Hybrid search (semantic + BM25 with RRF)
│
├── io/              # File I/O
│   ├── mod.rs
│   ├── reader.rs    # File reading with mmap
│   └── unicode.rs   # Unicode/grapheme utilities
│
└── cli/             # Command-line interface
    ├── mod.rs
    ├── parser.rs    # Clap argument definitions
    ├── commands.rs  # Command implementations
    └── output.rs    # Output formatting
```

## Core Types

### Buffer

Represents a loaded file with metadata:

```rust
pub struct Buffer {
    pub id: Option<i64>,
    pub name: String,
    pub content: String,
    pub source: Option<String>,
    pub metadata: BufferMetadata,
}

pub struct BufferMetadata {
    pub size: usize,
    pub line_count: usize,
    pub hash: String,
    pub content_type: Option<String>,
    pub chunk_count: usize,
    pub created_at: Option<String>,
    pub updated_at: Option<String>,
}
```

### Chunk

Represents a segment of buffer content:

```rust
pub struct Chunk {
    pub buffer_id: i64,
    pub content: String,
    pub byte_range: Range<usize>,
    pub index: usize,
    pub metadata: ChunkMetadata,
}

pub struct ChunkMetadata {
    pub token_count: Option<usize>,
    pub has_overlap: bool,
    pub strategy: Option<String>,
}
```

### Context

Manages variables and state:

```rust
pub struct Context {
    buffers: HashMap<i64, Buffer>,
    variables: HashMap<String, ContextValue>,
    globals: HashMap<String, ContextValue>,
}

pub enum ContextValue {
    String(String),
    Number(i64),
    Float(f64),
    Boolean(bool),
    List(Vec<ContextValue>),
}
```

## Chunking System

### Chunker Trait

All chunking strategies implement:

```rust
pub trait Chunker: Send + Sync {
    fn chunk(
        &self,
        buffer_id: i64,
        text: &str,
        metadata: Option<&ChunkMetadata>,
    ) -> Result<Vec<Chunk>>;

    fn name(&self) -> &'static str;
    fn supports_parallel(&self) -> bool;
    fn description(&self) -> &'static str;
    fn validate(&self, metadata: Option<&ChunkMetadata>) -> Result<()>;
}
```

### Strategy Selection

| Strategy | Algorithm | Use Case |
|----------|-----------|----------|
| `SemanticChunker` | Unicode sentence/paragraph boundaries | Markdown, prose |
| `CodeChunker` | Language-aware function/class boundaries | Source code files |
| `FixedChunker` | Character boundaries with UTF-8 safety | Logs, raw text |
| `ParallelChunker` | Rayon-parallelized fixed chunking | Large files (>10MB) |

### Code Chunker Languages

The `CodeChunker` uses regex-based pattern matching for multiple languages:

| Language | Extensions | Boundary Detection |
|----------|------------|-------------------|
| Rust | .rs | `fn`, `impl`, `struct`, `enum`, `mod` |
| Python | .py | `def`, `class`, `async def` |
| JavaScript/TypeScript | .js, .jsx, .ts, .tsx | `function`, `class`, `const =` |
| Go | .go | `func`, `type` |
| Java | .java | `class`, `interface`, method signatures |
| C/C++ | .c, .cpp, .h, .hpp | Function definitions |
| Ruby | .rb | `def`, `class`, `module` |
| PHP | .php | `function`, `class` |

### Default Configuration

```rust
pub const DEFAULT_CHUNK_SIZE: usize = 3_000;    // ~750 tokens
pub const DEFAULT_OVERLAP: usize = 500;          // Context continuity
pub const MAX_CHUNK_SIZE: usize = 50_000;        // Safety limit
```

## Storage Layer

### Storage Trait

```rust
pub trait Storage: Send + Sync {
    // Buffer operations
    fn add_buffer(&mut self, buffer: &Buffer) -> Result<i64>;
    fn get_buffer(&self, id: i64) -> Result<Option<Buffer>>;
    fn get_buffer_by_name(&self, name: &str) -> Result<Option<Buffer>>;
    fn update_buffer(&mut self, buffer: &Buffer) -> Result<()>;
    fn delete_buffer(&mut self, id: i64) -> Result<()>;
    fn list_buffers(&self) -> Result<Vec<Buffer>>;

    // Chunk operations
    fn add_chunks(&mut self, buffer_id: i64, chunks: &[Chunk]) -> Result<()>;
    fn get_chunks(&self, buffer_id: i64) -> Result<Vec<Chunk>>;
    fn delete_chunks(&mut self, buffer_id: i64) -> Result<()>;

    // Variable operations
    fn set_variable(&mut self, name: &str, value: &ContextValue) -> Result<()>;
    fn get_variable(&self, name: &str) -> Result<Option<ContextValue>>;
    fn delete_variable(&mut self, name: &str) -> Result<()>;

    // Global operations
    fn set_global(&mut self, name: &str, value: &ContextValue) -> Result<()>;
    fn get_global(&self, name: &str) -> Result<Option<ContextValue>>;
    fn delete_global(&mut self, name: &str) -> Result<()>;
}
```

### SQLite Schema

```sql
-- Buffers table
CREATE TABLE buffers (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    name TEXT NOT NULL UNIQUE,
    content TEXT NOT NULL,
    source TEXT,
    size INTEGER NOT NULL,
    line_count INTEGER NOT NULL,
    hash TEXT NOT NULL,
    content_type TEXT,
    chunk_count INTEGER DEFAULT 0,
    created_at TEXT DEFAULT CURRENT_TIMESTAMP,
    updated_at TEXT DEFAULT CURRENT_TIMESTAMP
);

-- Chunks table
CREATE TABLE chunks (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    buffer_id INTEGER NOT NULL REFERENCES buffers(id) ON DELETE CASCADE,
    content TEXT NOT NULL,
    byte_start INTEGER NOT NULL,
    byte_end INTEGER NOT NULL,
    chunk_index INTEGER NOT NULL,
    token_count INTEGER,
    has_overlap INTEGER DEFAULT 0,
    strategy TEXT,
    created_at TEXT DEFAULT CURRENT_TIMESTAMP
);

-- Variables table
CREATE TABLE variables (
    name TEXT PRIMARY KEY,
    value TEXT NOT NULL,
    value_type TEXT NOT NULL,
    created_at TEXT DEFAULT CURRENT_TIMESTAMP,
    updated_at TEXT DEFAULT CURRENT_TIMESTAMP
);

-- Globals table
CREATE TABLE globals (
    name TEXT PRIMARY KEY,
    value TEXT NOT NULL,
    value_type TEXT NOT NULL,
    created_at TEXT DEFAULT CURRENT_TIMESTAMP,
    updated_at TEXT DEFAULT CURRENT_TIMESTAMP
);
```

## I/O Layer

### Memory-Mapped File Reading

For large files, `memmap2` provides efficient reading:

```rust
pub fn read_file(path: &Path) -> Result<String> {
    let file = File::open(path)?;
    let mmap = unsafe { MmapOptions::new().map(&file)? };
    let content = std::str::from_utf8(&mmap)?;
    Ok(content.to_string())
}
```

### Unicode Handling

The `unicode-segmentation` crate ensures proper handling of:
- Multi-byte UTF-8 characters
- Grapheme clusters
- Sentence boundaries

```rust
pub const fn find_char_boundary(s: &str, pos: usize) -> usize {
    if pos >= s.len() {
        return s.len();
    }
    let bytes = s.as_bytes();
    let mut boundary = pos;
    // UTF-8 continuation bytes start with 10xxxxxx (0x80-0xBF)
    while boundary > 0 && (bytes[boundary] & 0xC0) == 0x80 {
        boundary -= 1;
    }
    boundary
}
```

## Error Handling

All errors use `thiserror` for ergonomic error types:

```rust
#[derive(Error, Debug)]
pub enum Error {
    #[error("Storage error: {0}")]
    Storage(#[from] StorageError),

    #[error("Chunking error: {0}")]
    Chunking(#[from] ChunkingError),

    #[error("I/O error: {0}")]
    Io(#[from] std::io::Error),

    #[error("Command error: {0}")]
    Command(#[from] CommandError),
}
```

## RLM Pattern Implementation

### Concept Mapping

| RLM Concept | rlm-rs Implementation |
|-------------|----------------------|
| Root LLM | Claude Code main conversation (Opus/Sonnet) |
| Sub-LLM | Claude Code subagent (Haiku) |
| External Environment | `rlm-rs` CLI + SQLite database |
| Chunk | `Chunk` struct with byte range and metadata |
| Buffer | `Buffer` struct with full content |
| State | SQLite persistence + context variables |

### Workflow

1. **Load**: Large document loaded into buffer, chunked, stored in SQLite
2. **Index**: Root LLM queries chunk indices via `chunk-indices`
3. **Process**: Sub-LLM processes individual chunks via file reads
4. **Aggregate**: Results stored back via `add-buffer`
5. **Synthesize**: Root LLM synthesizes final result

## Performance Considerations

### Token Estimation

Chunks target ~10,000 tokens to fit within Claude's 25,000 token read limit:

```rust
impl Chunk {
    pub fn estimate_tokens(&self) -> usize {
        // Approximate: 4 characters per token
        self.content.len() / 4
    }
}
```

### Parallel Processing

The `ParallelChunker` uses Rayon for multi-threaded chunking:

```rust
impl Chunker for ParallelChunker {
    fn chunk(&self, buffer_id: i64, text: &str, metadata: Option<&ChunkMetadata>) -> Result<Vec<Chunk>> {
        let segments = split_into_segments(text, self.segment_count);

        segments
            .par_iter()
            .enumerate()
            .flat_map(|(i, segment)| {
                self.inner.chunk(buffer_id, segment, metadata)
            })
            .collect()
    }
}
```

## Testing Strategy

### Unit Tests

Each module has `#[cfg(test)]` tests:

```rust
#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_chunk_boundaries() {
        let chunker = SemanticChunker::with_size(100);
        let chunks = chunker.chunk(1, "Hello. World.", None).unwrap();
        assert!(!chunks.is_empty());
    }
}
```

### Integration Tests

`tests/integration_test.rs` covers end-to-end workflows.

### Property-Based Tests

Using `proptest` for invariant verification:

```rust
proptest! {
    #[test]
    fn chunk_byte_range_valid(content in ".{1,1000}") {
        let chunker = FixedChunker::with_size(100);
        let chunks = chunker.chunk(1, &content, None).unwrap();
        for chunk in chunks {
            prop_assert!(chunk.byte_range.end <= content.len());
        }
    }
}
```

## Search System

### Hybrid Search Architecture

rlm-rs implements a hybrid search system combining multiple retrieval methods:

```
┌─────────────────────────────────────────────────────────────┐
│                      Search Query                            │
└─────────────────────────────────────────────────────────────┘
              ┌───────────────┼───────────────┐
              ▼               ▼               ▼
    ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
    │ Semantic Search │ │  BM25 Search    │ │  HNSW Index     │
    │  (Embeddings)   │ │  (FTS5)         │ │  (Optional)     │
    └────────┬────────┘ └────────┬────────┘ └────────┬────────┘
             │                   │                   │
             └───────────────────┼───────────────────┘
                    ┌───────────────────────┐
                    │  Reciprocal Rank      │
                    │  Fusion (RRF)         │
                    └───────────────────────┘
                    ┌───────────────────────┐
                    │   Ranked Results      │
                    └───────────────────────┘
```

### Embedding System

| Component | Implementation | Details |
|-----------|---------------|---------|
| Model | BGE-M3 via fastembed | 1024 dimensions |
| Fallback | Hash-based embedder | When fastembed unavailable |
| Storage | SQLite BLOB | Compact binary storage |
| Incremental | `embed_buffer_chunks_incremental` | Only new/changed chunks |

### HNSW Index (Optional)

When the `usearch-hnsw` feature is enabled:

- O(log n) approximate nearest neighbor search
- Persistent index on disk
- Incremental updates
- Falls back to brute-force when disabled

## Future Extensions

### Planned Features

- **Streaming**: Process chunks as they're generated
- **Compression**: Compress stored content
- **Encryption**: Encrypt sensitive buffers

### Extension Points

- `Chunker` trait for custom chunking strategies
- `Embedder` trait for alternative embedding models
- `Storage` trait for alternative backends (PostgreSQL, Redis)
- Output formatters for additional formats (YAML, TOML)

---

## See Also

- [RLM-Inspired Design]rlm-inspired-design.md - How rlm-rs builds on the RLM paper
- [Plugin Integration]plugin-integration.md - Claude Code plugin setup and portability
- [CLI Reference]cli-reference.md - Complete command documentation
- [API Reference]api.md - Rust library documentation
- [README.md]../README.md - Project overview
- [RLM Paper]https://arxiv.org/abs/2512.24601 - Original research paper