How to: chunk a CAD file for RAG

Goal: convert a CAD file into chunks you can embed and store in a vector database for retrieval-augmented generation.

This assumes you can already parse a file with cadling.

1. Parse, then chunk

from cadling import DocumentConverter
from cadling.chunker.hybrid_chunker import CADHybridChunker

doc = DocumentConverter().convert("assembly.step").document

chunker = CADHybridChunker(max_tokens=512, overlap_tokens=50)
chunks = list(chunker.chunk(doc))

Pick the chunker for your retrieval need:

Chunker	Strategy
CADHybridChunker	entity-level + semantic grouping (good default)
CADHierarchicalChunker	assembly-hierarchy-aware (preserves BOM structure)

2. Use the chunk text and metadata

Each chunk carries text plus structured metadata you should store alongside the vector — it makes retrieval filterable and the context richer.

for chunk in chunks:
    text = chunk.text                 # what you embed
    meta = chunk.meta                 # entity types, topology subgraph, bbox
    print(chunk.chunk_id, len(meta.entity_ids), "entities")

3. Embed and index

Use any embedding model and vector store. The pattern:

records = []
for chunk in chunks:
    records.append({
        "id": chunk.chunk_id,
        "text": chunk.text,
        "vector": embed(chunk.text),          # your embedding function
        "metadata": {
            "entity_ids": list(chunk.meta.entity_ids),
            "source": "assembly.step",
        },
    })

vector_store.upsert(records)                  # your vector DB client

Store the metadata so you can filter retrieval (e.g. by entity type) and cite the source CAD file in answers.

4. From the CLI instead

cadling chunk assembly.step --max-tokens 512 --overlap 50 -o chunks.jsonl

Each line of chunks.jsonl is one chunk with its text and metadata, ready to feed an embedding/indexing job.

See also

• cadling Usage — all chunkers and the SDG pipeline.
• Generate synthetic Q&A to build training data from the same documents.