Building code-chunk: AST Aware Code Chunking

At Supermemory, we're building context engineering infrastructure for AI. A huge part of that is dealing with code: ingesting repos, understanding structure, and making it searchable. The problem is that most code chunking solutions are terrible.

We built code-chunk to fix this. It's now the best AST-based code chunking library out there, and I want to walk through how we built it.

The Problem with Text Splitters

Here's what most RAG pipelines do with code:

Source code → Split by character count → Generate embeddings → Store in vector DB

This is fine for prose. But code isn't prose. When you split Python at 500 characters, you end up with chunks like:

def calculate_total(items):
    """Calculate the total price of all items with tax."""
    subtotal = 0
    for item in items:
        subtotal += item.price * item.qu

Congratulations, you just cut a function in half. The embedding for this chunk has no idea what qu is. It doesn't know the function returns anything. If someone searches for "calculate total with tax", they might get this useless fragment. Text splitters don't understand code structure.

Enter cAST

I found this paper from CMU called "cAST: Enhancing Code Retrieval-Augmented Generation with Structural Chunking via Abstract Syntax Tree". The core idea is elegant:

1. Parse code into an AST (Abstract Syntax Tree)
2. Use the tree structure to find natural split points (functions, classes, methods)
3. Chunk at semantic boundaries instead of arbitrary character limits

The paper shows solid improvements: 4.3 points on Recall@5 for RepoEval, 2.67 points on Pass@1 for SWE-bench. But reading through the paper, I realized their implementation was missing a lot of what you actually need in production.

So we took the core idea and built something better.

How code-chunk Works

Step 1: Parse with tree-sitter

We use tree-sitter to parse source code into an AST. tree-sitter is battle-tested (it powers syntax highlighting in editors like Neovim, Helix, and Zed) and supports basically every language.

import { chunk } from 'code-chunk'

const chunks = await chunk('src/user.ts', sourceCode)

From a filepath, we detect the language and load the appropriate tree-sitter grammar. The result is a structured tree where every node has a type (function_declaration, class_body, import_statement, etc.) and byte ranges.

Step 2: Extract Entities

Here's where we diverge from the paper. We don't just traverse the tree. We extract semantic entities:

interface ExtractedEntity {
  type: 'function' | 'method' | 'class' | 'interface' | 'type' | 'enum' | 'import'
  name: string
  signature: string              // e.g., "async getUser(id: string): Promise<User>"
  docstring: string | null       // JSDoc, docstrings, etc.
  byteRange: { start: number; end: number }
  lineRange: { start: number; end: number }
  parent: string | null          // Which class/function contains this
}

For each entity, we extract:

• The full signature (not just the name)
• Any documentation comments
• Parent relationships (a method knows it belongs to a class)
• Import sources and symbols

This metadata becomes critical later.

Step 3: Build the Scope Tree

Entities are organized into a hierarchical scope tree. A method inside a class knows its parent. A nested function knows its containing function.

interface ScopeNode {
  entity: ExtractedEntity
  children: ScopeNode[]
  parent: ScopeNode | null
}

We build this by sorting entities by byte range and using DFS to find the deepest container for each entity:

const findParentNode = (roots: ScopeNode[], entity: ExtractedEntity): ScopeNode | null => {
  const findInNode = (node: ScopeNode): ScopeNode | null => {
    // Does this node contain the entity?
    if (!rangeContains(node.entity.byteRange, entity.byteRange)) {
      return null
    }
    // Check children for deeper match
    for (const child of node.children) {
      const deeperMatch = findInNode(child)
      if (deeperMatch) return deeperMatch
    }
    // No child contains it, so this is the deepest
    return node
  }
  
  for (const root of roots) {
    const found = findInNode(root)
    if (found) return found
  }
  return null
}

This scope tree lets us answer questions like "where does this code live?" with context like UserService > getUser.

Step 4: The Chunking Algorithm

The paper calls it "recursive split-then-merge". The idea is to pack as many complete syntactic units as possible into each chunk without exceeding a size limit. We do this with greedy window assignment:

function* greedyAssignWindows(nodes: SyntaxNode[], code: string, maxSize: number) {
  let currentWindow = { nodes: [], size: 0 }
  
  for (const node of nodes) {
    const nodeSize = getNwsCount(node) // Non-whitespace character count
    
    if (currentWindow.size + nodeSize <= maxSize) {
      // Node fits - add it
      currentWindow.nodes.push(node)
      currentWindow.size += nodeSize
    } else if (nodeSize > maxSize) {
      // Node is too big - recurse into children
      yield currentWindow
      yield* greedyAssignWindows(node.children, code, maxSize)
      currentWindow = { nodes: [], size: 0 }
    } else {
      // Node doesn't fit but isn't oversized - start new window
      yield currentWindow
      currentWindow = { nodes: [node], size: nodeSize }
    }
  }
  
  if (currentWindow.nodes.length > 0) {
    yield currentWindow
  }
}

We iterate through AST nodes and keep adding them to the current window until we hit the size limit. If a node is too big to fit anywhere, we recurse into its children and chunk those instead. If a node just doesn't fit in the current window but isn't oversized, we start a new window.

The paper uses non-whitespace character count instead of lines, and we do the same. Two chunks with the same line count can have wildly different amounts of actual code. A file full of blank lines and comments shouldn't count the same as dense logic.

We precompute a cumulative sum array for O(1) range queries:

const preprocessNwsCumsum = (code: string): Uint32Array => {
  const cumsum = new Uint32Array(code.length + 1)
  let count = 0
  for (let i = 0; i < code.length; i++) {
    if (code.charCodeAt(i) > 32) count++ // Not whitespace
    cumsum[i + 1] = count
  }
  return cumsum
}

// O(1) range query
const getNwsCount = (start: number, end: number) => cumsum[end] - cumsum[start]

After generating windows, we merge adjacent small ones to reduce fragmentation:

function* mergeAdjacentWindows(windows: Generator<ASTWindow>, maxSize: number) {
  let current: ASTWindow | null = null
  
  for (const window of windows) {
    if (!current) {
      current = window
    } else if (current.size + window.size <= maxSize) {
      current = mergeWindows(current, window)
    } else {
      yield current
      current = window
    }
  }
  
  if (current) yield current
}

Step 5: Contextualized Text

Here's where code-chunk really differentiates itself. Raw chunk text isn't enough for good embeddings. You need context.

Each chunk includes a contextualizedText field:

# src/services/user.ts
# Scope: UserService
# Defines: async getUser(id: string): Promise<User>
# Uses: Database
# After: constructor

  async getUser(id: string): Promise<User> {
    return this.db.query('SELECT * FROM users WHERE id = ?', [id])
  }

This prepends semantic context to the raw code:

1. File path: where this code lives
2. Scope chain: what class/function contains it
3. Entity signatures: what's defined here
4. Imports used: dependencies
5. Sibling context: what comes before/after

Why does this matter? Embedding models are trained on natural language. When you embed async getUser(id: string), the model doesn't inherently know this is inside a UserService class or that it uses a Database. By prepending this context, the embedding captures semantic relationships that pure code misses.

The implementation:

function formatChunkWithContext(text: string, context: ChunkContext): string {
  const parts: string[] = []
  
  if (context.filepath) {
    parts.push(`# ${context.filepath.split('/').slice(-3).join('/')}`)
  }
  
  if (context.scope.length > 0) {
    parts.push(`# Scope: ${context.scope.map(s => s.name).reverse().join(' > ')}`)
  }
  
  const signatures = context.entities
    .filter(e => e.signature && e.type !== 'import')
    .map(e => e.signature)
  if (signatures.length > 0) {
    parts.push(`# Defines: ${signatures.join(', ')}`)
  }
  
  if (context.imports.length > 0) {
    parts.push(`# Uses: ${context.imports.slice(0, 10).map(i => i.name).join(', ')}`)
  }
  
  parts.push('')
  parts.push(text)
  return parts.join('\n')
}

What We Added Beyond the Paper

The cAST paper is a great foundation, but it's research code. Here's what we built on top:

1. Rich context extraction. The paper just chunks. We extract full entity metadata, build scope trees, and format context for embeddings.

2. Overlap support. Chunks can include the last N lines from the previous chunk. This helps with queries that target code at chunk boundaries.

const chunks = await chunk('file.ts', code, { overlapLines: 10 })

3. Streaming. Process large files without loading everything into memory:

for await (const chunk of chunkStream('large-file.ts', code)) {
  await process(chunk)
}

4. Batch processing. Chunk entire codebases with controlled concurrency:

const results = await chunkBatch(files, { 
  concurrency: 10,
  onProgress: (done, total) => console.log(`${done}/${total}`)
})

5. Effect integration. First-class support for the Effect library:

const program = Stream.runForEach(
  chunkStreamEffect('file.ts', code),
  (chunk) => Effect.log(chunk.text)
)

6. WASM support. Works in Cloudflare Workers and other edge runtimes:

import { createChunkerFromWasm } from 'code-chunk/wasm'
import treeSitterWasm from 'web-tree-sitter/tree-sitter.wasm'
import typescriptWasm from 'tree-sitter-typescript/tree-sitter-tsx.wasm'

const chunker = await createChunkerFromWasm({
  treeSitter: treeSitterWasm,
  languages: { typescript: typescriptWasm }
})

Evaluating It

We built an eval harness using RepoEval data to measure how well this actually works. First run: 100% recall across all chunkers. That's not a good result. That's a broken benchmark.

The Benchmark Was Too Easy

The original setup had a fundamental flaw. The query was a code prefix (say, lines 0-19 of a file), and the target was the same file containing that exact prefix. Embedding models trivially found the target because the query text was literally inside the target file.

This is like asking someone to find a book when you've already told them the exact title and they're standing in front of it.

Fixing the Eval

First, we added hard negatives: 500 distractor files from the same repository. Same-repo files share vocabulary and coding style, making them genuinely hard to distinguish from the target. Before we had 108 files from the target repo. After, we had the target file plus 500 same-repo distractors.

Second, we added an Intersection over Union (IoU) threshold of 0.3. A chunk is only "relevant" if it actually overlaps with the ground truth lines, not just somewhere in the same file.

Ground truth: lines 0-19 (20 lines)

Chunk A (lines 0-61):  IoU = 20/62  = 0.32 ✓ Relevant
Chunk B (lines 0-100): IoU = 20/101 = 0.20 ✗ Too bloated
Chunk C (lines 0-25):  IoU = 20/26  = 0.77 ✓ Excellent

This penalizes chunkers that create oversized chunks. Getting "somewhere in the file" isn't good enough. You need to actually hit the relevant code.

We also started tracking IoU directly as a quality metric. Recall is binary: did you find it or not. IoU is continuous: how well did you find it. We track iou_at_5 and iou_at_10, the average best IoU score in top-k retrieved chunks.

Results

After fixing the benchmark:

Chunker

Recall@5

IoU@5

code-chunk (AST)

70.1%

0.43

chonkie-code

49.0%

0.38

Fixed-size baseline

42.4%

0.34

The AST-based chunker creates chunks that align better with semantic boundaries. Higher IoU means the retrieved chunks actually contain the relevant code, not just code from the same file.

We also ran an agent-based eval on SWE-bench Lite comparing ops-only (Read/Grep/Glob) vs ops+semantic-search:

Duration

Tokens

Cost

Tool Calls

Without search

2.0m

4.3k

$0.25

19

With semantic search

1.2m

2.4k

$0.20

12

The semantic search agent finds relevant files faster, especially for queries that need code understanding rather than just text matching.

Using code-chunk

Installation:

bun add code-chunk
# or npm install code-chunk

Basic usage:

import { chunk } from 'code-chunk'

const chunks = await chunk('src/user.ts', sourceCode)

for (const c of chunks) {
  console.log(c.text)                // Raw code
  console.log(c.contextualizedText)  // Code with context for embeddings
  console.log(c.context.scope)       // [{ name: 'UserService', type: 'class' }]
  console.log(c.context.entities)    // [{ name: 'getUser', type: 'method', ... }]
}

For RAG, use contextualizedText when generating embeddings:

for (const c of chunks) {
  const embedding = await embed(c.contextualizedText)
  await vectorDB.upsert({
    id: `${filepath}:${c.index}`,
    embedding,
    metadata: { filepath, lines: c.lineRange }
  })
}

Supported languages: TypeScript, JavaScript, Python, Rust, Go, Java.

Wrapping Up

Chunking code well means more than respecting size limits. You need to preserve and enrich semantic context, and that's what code-chunk does.

If you're building code RAG and using naive text splitters, you're leaving performance on the table. Give code-chunk a try.

• GitHub: supermemoryai/code-chunk
• npm: code-chunk
• Paper: cAST: Enhancing Code Retrieval-Augmented Generation

Questions? I'm on Twitter/X.