Swarmauri Parser Bert Embedding

swarmauri_parser_bertembedding is the Swarmauri embedding parser for turning text into dense vector representations with Hugging Face BERT models. It returns Swarmauri Document objects whose content keeps the original text and whose metadata stores the generated embedding vector.

Why Use Swarmauri Parser Bert Embedding

• Generate dense semantic vectors inside a Swarmauri parser-style workflow.
• Keep original text and embedding output together in a single Document object.
• Swap BERT model names when you need a different encoder surface.
• Feed embeddings into retrieval, clustering, semantic search, reranking, or downstream vector store pipelines.

FAQ

What does this parser output?
Swarmauri Document objects containing the original text and an averaged BERT embedding stored in metadata["embedding"].

What model does it use by default?
bert-base-uncased.

Can it parse a batch of strings?
Yes. The current implementation accepts a single string or a list of strings.

Does it download model weights?
Yes. On first use, Hugging Face model and tokenizer assets are downloaded if they are not already cached locally.

Features

• Dense embedding generation using BertTokenizer and BertModel.
• Supports single-string and batch-text parsing.
• Stores the original text alongside the embedding vector in each document.
• Uses inference mode with torch.no_grad() and mean token pooling.
• Supports Python 3.10, 3.11, 3.12, 3.13, and 3.14.

Installation

uv add swarmauri_parser_bertembedding

pip install swarmauri_parser_bertembedding

Notes:

• First-run model downloads come from Hugging Face.
• Install a CUDA-enabled PyTorch build separately if GPU execution is required.

Usage

from swarmauri_parser_bertembedding import BERTEmbeddingParser

parser = BERTEmbeddingParser(parser_model_name="bert-base-uncased")
documents = parser.parse(
    [
        "Swarmauri agents cooperate over shared memory.",
        "Dense embeddings power semantic search.",
    ]
)

for document in documents:
    embedding = document.metadata["embedding"]
    print(document.content)
    print(len(embedding), embedding[:5])

Examples

Embed a single sentence

from swarmauri_parser_bertembedding import BERTEmbeddingParser

parser = BERTEmbeddingParser()
documents = parser.parse("Composable intelligence infrastructure")

print(documents[0].id)
print(documents[0].metadata["source"])
print(documents[0].metadata["embedding"].shape)

Embed a batch for downstream storage

from swarmauri_parser_bertembedding import BERTEmbeddingParser

texts = [
    "Customer support workflows need retrieval.",
    "Embeddings support semantic matching.",
    "Vector stores preserve nearest-neighbor search state.",
]

parser = BERTEmbeddingParser()
documents = parser.parse(texts)

for document in documents:
    vector = document.metadata["embedding"]
    print(document.id, len(vector))

Use an alternate BERT model name

from swarmauri_parser_bertembedding import BERTEmbeddingParser

parser = BERTEmbeddingParser(parser_model_name="bert-base-multilingual-cased")
docs = parser.parse("Bonjour tout le monde")
print(docs[0].metadata["embedding"][:5])

Related Packages

• swarmauri_embedding_mlm
• swarmauri_vectorstore_mlm
• swarmauri_vectorstore_qdrant
• swarmauri_vectorstore_pinecone

Swarmauri Foundations

• swarmauri
• swarmauri_core
• swarmauri_base
• swarmauri_standard

More Documentation

• Hugging Face Transformers BERT docs
• Transformers AutoTokenizer and tokenization docs
• PyTorch documentation

Best Practices

• Chunk very long texts before parsing so they stay within the BERT token limit.
• Cache Hugging Face assets in CI and deployment environments to avoid repeated model downloads.
• Use a model variant aligned to your language and domain.
• Persist vectors into a Swarmauri vector store if you plan to search or reuse them beyond a single process.

License

This project is licensed under the Apache-2.0 License.