Is it possible to use aws bedrock embeddings or Custom Embedding for HybridChunker?

[princeGedeon]

For process complex document with tables and unstructured table in AWS ecosystem, can i use Docling with aws embeddings via Bedrock?

[dosubot[bot]]

Yes, you can use Docling with AWS Bedrock embeddings, but there's an important distinction to understand:

HybridChunker only handles chunking, not embedding generation. It uses a tokenizer to determine chunk boundaries (counting tokens to stay within limits), but outputs text chunks—not embeddings. The actual embedding step happens separately in your RAG pipeline.

For AWS Bedrock integration, you have two approaches:

1. Use HybridChunker for chunking, then embed with Bedrock separately:

   from docling_core.transforms.chunker import HybridChunker
   
   chunker = HybridChunker(max_tokens=512)
   chunks = list(chunker.chunk(doc))
   
   # Then embed each chunk using your Bedrock client
   for chunk in chunks:
       embedding = bedrock_client.embed(chunk.text)

2. Create a custom tokenizer for Bedrock models by implementing the BaseTokenizer interface:

   from docling_core.transforms.chunker.tokenizer.base import BaseTokenizer
   
   class BedrockTokenizer(BaseTokenizer):
       def count_tokens(self, text: str) -> int:
           # Implement token counting for your Bedrock model
           pass
       def get_max_tokens(self) -> int:
           return 8192  # or your model's limit
       def get_tokenizer(self):
           return self

Built-in tokenizers include HuggingFace and OpenAI tiktoken. HybridChunker also handles complex tables well, automatically splitting large tables across chunks while preserving headers.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}