Silas-Embedding

Model Description

Silas-Embedding is an embedding model from Convence Lab for semantic search, retrieval, reranking pipelines, and multimodal embedding experiments.

The model is designed to embed queries and documents into a shared vector space for fast retrieval. It can be used for document search, question-to-passage matching, image-text retrieval experiments, and retrieval-augmented generation pipelines.

• Developed by: Convence Lab
• Model name: Silas-Embedding
• Model type: Embedding model
• Primary task: Text and multimodal retrieval
• License: Apache 2.0

---

Core Capabilities

Silas-Embedding focuses on practical retrieval behavior:

• Semantic Search - Match natural-language queries to relevant passages or documents.
• Document Retrieval - Retrieve policy notes, memos, document chunks, and structured text.
• RAG Pipelines - Use embeddings as the retrieval layer before generation.
• Similarity Scoring - Compare query/document or sentence-pair similarity.
• Multimodal Retrieval Experiments - Supports image-text style embedding workflows when used with compatible tooling.

---

Benchmark Results

ParseEmbed

ParseEmbed is an internal Convence retrieval benchmark candidate. These results are provided as an early, unverified benchmark score while the benchmark is still being prepared for broader validation.

Benchmark	Split	Queries	Corpus	Recall@1	Recall@5	Recall@10	MRR@10	NDCG@10
ParseEmbed	test	720	2,880	51.67%	98.06%	99.86%	72.71%	79.61%

Evaluation settings:

• Dataset: Convence/ParseEmbed
• Query config: queries
• Corpus config: corpus
• Qrels config: default
• Max sequence length: 512
• Batch size: 32
• Evaluation date: 2026-05-21

---

Getting Started

Install dependencies:

pip install -U sentence-transformers torch torchvision transformers

Load the model:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Convence/Silas-Embedding", trust_remote_code=True)

queries = [
    "Find the memo where payment hold is capped at 96 hours during Q1.",
]

documents = [
    "Cedar privacy requests memo: The payment hold is capped at 96 hours during Q1, after the second failed retry.",
    "Cedar privacy requests memo: The payment hold target is at least 96 hours during Q1.",
]

query_embeddings = model.encode(queries, normalize_embeddings=True)
document_embeddings = model.encode(documents, normalize_embeddings=True)

scores = query_embeddings @ document_embeddings.T
print(scores)

For retrieval, use the highest-scoring documents as candidates:

import numpy as np

top_k = 2
ranking = np.argsort(-scores[0])[:top_k]

for idx in ranking:
    print(float(scores[0][idx]), documents[idx])

---

Recommended Usage

Semantic Retrieval

Use short instruction prefixes for better retrieval consistency:

query: <user question>
passage: <document text>

Example:

query = "query: Find the refund policy for enterprise customers."
passage = "passage: Enterprise refund requests must be reviewed within 7 business days."

RAG

Recommended retrieval flow:

1. Split documents into chunks.
2. Embed chunks with Silas-Embedding.
3. Store vectors in a vector database.
4. Embed the user query.
5. Retrieve top-k chunks.
6. Optionally rerank the top results with a reranker.
7. Pass the final context to a language model.

ParseEmbed-Style Retrieval

For high-precision first-result retrieval, pair Silas-Embedding with a reranker. The ParseEmbed result shows strong candidate retrieval at Recall@5 and Recall@10, while Recall@1 can still improve with more hard-negative training.

---

Model Details

Property	Silas-Embedding
Organization	Convence Lab
Primary Task	Embedding / retrieval
Supported Use Cases	Semantic search, RAG, document retrieval, similarity scoring
Recommended Max Sequence Length	512
Library	Sentence Transformers
License	Apache 2.0

---

Limitations

• ParseEmbed is currently an internal benchmark candidate and should not be treated as an official leaderboard result yet.
• Top-1 retrieval is still improving; reranking is recommended for high-stakes retrieval.
• Embeddings may be sensitive to prompt formatting, chunk size, and document noise.
• Long documents should be chunked before embedding.
• The model may retrieve semantically similar but factually incorrect distractors when the corpus contains hard negatives.
• Do not use retrieval results without validation for legal, medical, financial, identity, or safety-critical decisions.

---

Ethics and Safety

Embedding models can be used to retrieve sensitive or private information from large document collections. Users are responsible for applying access controls, dataset permissions, privacy review, and logging policies when deploying this model.

Silas-Embedding should be used with care when indexing personal data, private records, confidential documents, or regulated information.

---

Citation

@misc{convence2026silasembedding,
  title={Silas-Embedding},
  author={Convence Lab},
  year={2026},
  url={https://huggingface.co/Convence/Silas-Embedding}
}