A specialized sentence-transformers model fine-tuned for semantic search and information retrieval in technical documentation, with a focus on enterprise infrastructure and virtualization technologies.

Model Details

Description

This model extends BAAI/bge-base-en-v1.5 with domain-specific fine-tuning for technical documentation retrieval. It generates 768-dimensional dense embeddings optimized for semantic similarity in enterprise technology contexts.

Model Type: Sentence Transformer (BERT-based)
Base Model: BAAI/bge-base-en-v1.5
Embedding Dimension: 768
Max Sequence Length: 512 tokens
Language: English
License: MIT

Intended Use

Primary Use Cases:

Semantic search over technical documentation
Information retrieval for enterprise infrastructure queries
RAG (Retrieval-Augmented Generation) pipelines
Technical support knowledge bases
Enterprise search systems

Optimized For:

Natural language queries about technical topics
Documentation retrieval and ranking
Question answering systems
Knowledge management platforms

Out-of-Scope

This model is specialized for technical documentation and may not perform optimally for:

General domain text
Non-English languages
Code search or generation
Creative writing or entertainment content

Quick Start

Installation

pip install sentence-transformers

Basic Usage

python

from sentence_transformers import SentenceTransformer, util

# Load model
model = SentenceTransformer('BarraHome/vmware-embeddings-large-v1')

# Example queries and documents
queries = [
    "How to configure high availability?",
    "Steps to install guest tools"
]

documents = [
    "High availability can be configured through the management interface...",
    "To install guest tools, first mount the ISO image..."
]

# Generate embeddings
query_embeddings = model.encode(queries)
doc_embeddings = model.encode(documents)

# Calculate similarity
similarities = util.cos_sim(query_embeddings, doc_embeddings)
print(similarities)

Semantic Search Example

python

from sentence_transformers import SentenceTransformer, util

model = SentenceTransformer('BarraHome/vmware-embeddings-large-v1')

# Your document corpus
corpus = [
    "Documentation about high availability features...",
    "Guide for load balancing configuration...",
    "Instructions for live migration procedures..."
]

# Encode corpus
corpus_embeddings = model.encode(corpus, convert_to_tensor=True)

# Query
query = "How to enable high availability?"
query_embedding = model.encode(query, convert_to_tensor=True)

# Search
hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=3)

# Display results
for hit in hits[0]:
    print(f"Score: {hit['score']:.4f}")
    print(f"Document: {corpus[hit['corpus_id']]}\n")

Performance

Evaluation Metrics

Evaluated on a held-out test set of 2,000 diverse technical queries:

Metric	Base Model	Fine-tuned	Improvement
Recall@1	0.637	0.759	+19.2%
Recall@3	0.805	0.927	+15.2%
Recall@5	0.853	0.956	+12.1%
Recall@10	0.906	0.979	+8.0%
NDCG@10	0.775	0.879	+13.4%

Key Performance Indicators

✅ 75.9% top-1 accuracy
✅ 92.7% top-3 recall
✅ 97.9% top-10 recall
✅ 0.879 NDCG@10 (excellent ranking quality)

Comparison with Base Model

The fine-tuned model shows consistent improvements across all metrics:

Higher recall at all k values
Better ranking quality (NDCG)
More accurate top-1 predictions

Performance Visualizations

Detailed Metric Comparison:

Percentage Improvements:

Training Details

Training Configuration

Framework: sentence-transformers
Loss Function: MultipleNegativesRankingLoss
Training Strategy: Contrastive learning with hard negative mining
Epochs: 1
Batch Size: 64
Learning Rate: 2e-5 (with 10% warmup)
Training Samples: 671,972 query-document pairs
Total Steps: 10,500
Training Duration: 4 hours 6 minutes
Throughput: 45.4 samples/second
Final Loss: 2.245
Precision: FP16
Hardware: NVIDIA RTX A6000 (49GB VRAM)

Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': True})
  (1): Pooling({'pooling_mode_cls_token': True})
  (2): Normalize()
)

Limitations

Known Limitations

Domain-Specific: Optimized for technical documentation; general domain performance not guaranteed
English Only: No multi-language support
Context Length: Limited to 512 tokens
Recency: Knowledge current as of training date

Recommendations

For optimal results:

Query Formulation:
- Use natural language questions
- Include relevant technical terms
- Keep queries under 512 tokens
Hybrid Search:
- Combine with keyword search (BM25) for best results
- Use semantic search for understanding, keyword for precision
Batch Processing:
- Use encode(..., batch_size=32) for large collections
- Enable convert_to_tensor=True for GPU acceleration
Reranking:
- Consider using a cross-encoder for final reranking
- Retrieve top-100 with this model, rerank to top-10

Technical Specifications

Model Information

Parameters: ~110M
Architecture: BERT-base
Pooling: CLS token
Normalization: L2
Similarity Function: Cosine similarity

Performance Benchmarks

Hardware	Batch Size	Throughput
RTX 3090	32	~850 docs/sec
A100	128	~2,100 docs/sec
CPU (16 cores)	8	~180 docs/sec

Resource Requirements

Minimum:

GPU: 4GB VRAM (batch size 16)
CPU: 4 cores, 8GB RAM

Recommended:

GPU: 8GB+ VRAM (batch size 32+)
CPU: 8+ cores, 16GB+ RAM

Citation

bibtex

@misc{vmware-embeddings-2024,
  author = {Alberto Ferrer},
  title = {VMware Technical Documentation Embeddings},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/BarraHome/vmware-embeddings-large-v1}}
}

Base Model Citation

bibtex

@misc{bge-base-en-v1.5,
  author = {BAAI},
  title = {BGE Base English v1.5},
  year = {2023},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/BAAI/bge-base-en-v1.5}}
}

Acknowledgments

Base Model: BAAI/bge-base-en-v1.5 by Beijing Academy of Artificial Intelligence
Framework: sentence-transformers by UKPLab

License

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Note: This model is intended for research and development. For production use, ensure compliance with your organization's policies and applicable regulations.