README.md

---
license: apache-2.0
base_model:
- Qwen/Qwen3-Embedding-0.6B
pipeline_tag: feature-extraction
library_name: sentence-transformers
language:
- multilingual
tags:
- text-embeddings
- feature-extraction
- sentence-transformers
- transformers
- multilingual
- qwen3
- alfotech
- silas
- rag
- semantic-search
- vector-search
model-index:
- name: silas-embedding-0.6b
  results:
  - task:
      type: feature-extraction
      name: Embedding Model — Easy Benchmark
    metrics:
    - type: recall_at_1
      value: 0.967
      name: Recall@1
    - type: recall_at_5
      value: 0.999
      name: Recall@5
    - type: recall_at_10
      value: 1.0
      name: Recall@10
    - type: mrr_at_10
      value: 0.9815
      name: MRR@10
  - task:
      type: feature-extraction
      name: Embedding Model — Hard Benchmark
    metrics:
    - type: recall_at_1
      value: 0.614
      name: Recall@1
    - type: recall_at_5
      value: 0.883
      name: Recall@5
    - type: recall_at_10
      value: 0.929
      name: Recall@10
    - type: mrr_at_10
      value: 0.7268
      name: MRR@10
---

<p align="center">
  <img src="image.png" width="220"/>
</p>

<h1 align="center">Silas Embedding 0.6B</h1>

<p align="center">
  <b>Production-Grade Embedding Model by Alfo Tech Industries</b><br/>
  Multilingual · Long Context · Semantic Understanding · AI Infrastructure
</p>

<p align="center">
  <a href="https://huggingface.co/alfotech/silas-embedding-0.6b">
    <img src="https://img.shields.io/badge/🤗_Hugging_Face-Model-yellow" alt="Hugging Face"/>
  </a>
  <img src="https://img.shields.io/badge/Parameters-0.6B-6D5BFF" alt="Parameters"/>
  <img src="https://img.shields.io/badge/Context-32K-00E5FF" alt="Context"/>
  <img src="https://img.shields.io/badge/Embedding-1024D-8B5CF6" alt="Embedding Dimension"/>
  <img src="https://img.shields.io/badge/Languages-100%2B-22C55E" alt="Languages"/>
  <img src="https://img.shields.io/badge/License-Apache--2.0-2563EB" alt="License"/>
  <img src="https://img.shields.io/badge/Status-Active_Development-F59E0B" alt="Status"/>
</p>

<p align="center">
  <a href="#quick-start">Quick Start</a> ·
  <a href="#benchmark-dashboard">Benchmarks</a> ·
  <a href="#production-deployment">Deployment</a> ·
  <a href="#intended-use">Intended Use</a> ·
  <a href="#faq">FAQ</a> ·
  <a href="#citation">Citation</a>
</p>

---

## Table of Contents

- [Model Profile](#model-profile)
- [Why Silas?](#why-silas)
- [Embedding Capabilities](#embedding-capabilities)
- [Benchmark Dashboard](#benchmark-dashboard)
  - [Easy Embedding Benchmark](#easy-embedding-benchmark)
  - [Hard Embedding Benchmark](#hard-embedding-benchmark)
  - [Benchmark Comparison](#benchmark-comparison)
  - [Semantic Analysis](#semantic-analysis)
- [Development Architecture](#development-architecture)
- [Quick Start](#quick-start)
- [Usage Guide](#usage-guide)
  - [Query vs. Document Embeddings](#query-vs-document-embeddings)
  - [Variable Embedding Size (Matryoshka)](#variable-embedding-size-matryoshka)
  - [Batch Processing](#batch-processing)
- [Production Deployment](#production-deployment)
  - [Recommended Production Config](#production-starting-point)
  - [Serving Options](#serving-options)
  - [Integration Architecture](#integration)
- [Intended Use](#intended-use)
- [Bias, Risks & Limitations](#bias-risks--limitations)
- [Current Development Status](#current-development-status)
- [Roadmap](#roadmap)
- [Versioning](#versioning)
- [FAQ](#faq)
- [Reproducibility](#reproducibility)
- [Citation](#citation)
- [License](#license)
- [Contact & Support](#contact--support)

---

# Model Profile

**Silas Embedding 0.6B** is a production-oriented multilingual embedding model developed by **Alfo Tech Industries**.

Silas converts text into dense vector representations designed for semantic understanding, similarity measurement, knowledge systems, AI search, Retrieval-Augmented Generation, and vector databases. The model is based on **Qwen3-Embedding-0.6B** and follows an evaluation-first engineering workflow focused on difficult semantic matching scenarios rather than only easy, well-separated cases.

| Property                   | Specification         |
| --------------------------- | ---------------------- |
| **Model**                   | Silas Embedding 0.6B  |
| **Organization**            | Alfo Tech Industries  |
| **Base Model**               | Qwen3-Embedding-0.6B  |
| **Parameters**               | 0.6B                  |
| **Context Window**           | 32K                   |
| **Default Embedding Size**   | 1024 dimensions       |
| **Supported Dimensions**     | 32–1024               |
| **Languages**                | 100+                  |
| **Architecture Type**        | Text Embedding        |
| **Instruction Aware**        | Yes                   |
| **Framework**                | Sentence Transformers |
| **License**                  | Apache-2.0            |

---

# Why Silas?

A useful embedding model should do more than place related sentences close together — it should help a downstream system distinguish the *correct* semantic match from close, plausible-looking alternatives.

```text
Correct Semantic Meaning
            │
            ▼
     Silas Embedding
            │
            ▼
Dense Vector Representation
            │
     ┌──────┼────────┐
     ▼      ▼        ▼
   Search   RAG    Similarity
```

For this reason, Silas is evaluated on both a straightforward **Easy Benchmark** and a deliberately adversarial **Hard Benchmark** built from semantically similar candidates — the gap between the two is treated as the real signal of embedding quality, not the easy score alone.

---

# Embedding Capabilities

| Capability             | Example                                   |
| ----------------------- | ------------------------------------------ |
| **Semantic Search**      | Find conceptually relevant documents      |
| **RAG**                  | Retrieve context for language models      |
| **Knowledge Bases**      | Search enterprise documentation           |
| **FAQ Matching**         | Match questions with answers              |
| **Similarity Systems**   | Compare semantic meaning                  |
| **Code Search**          | Retrieve related programming content      |
| **Vector Databases**     | Store and query dense vectors             |
| **Multilingual AI**      | Represent text across supported languages |

---

# Benchmark Dashboard

## Evaluation Overview

| Benchmark                    | Corpus            | Queries | Primary Purpose                   |
| ------------------------------ | ------------------ | ------- | ----------------------------------- |
| **Easy Embedding Benchmark**   | Evaluation set     | 1,000   | Standard semantic matching        |
| **Hard Embedding Benchmark**   | 72,635 documents   | 1,000   | Difficult semantic discrimination |

The hard benchmark introduces semantically similar candidates, making it a more demanding test of embedding quality than corpus-level recall alone.

## Easy Embedding Benchmark

| Metric        | Score   |
| -------------- | ------- |
| **Recall@1**   | 96.70%  |
| **Recall@5**   | 99.90%  |
| **Recall@10**  | 100.00% |
| **MRR@10**     | 98.15%  |

**Interpretation:** The model performs strongly when the correct semantic match is relatively distinguishable from competing candidates — this reflects typical FAQ-matching and coarse retrieval workloads.

## Hard Embedding Benchmark

### Evaluation Setup

| Parameter                 | Value                           |
| --------------------------- | -------------------------------- |
| **Corpus Size**              | 72,635 documents                |
| **Evaluation Queries**       | 1,000                           |
| **Candidate Environment**    | Semantically similar documents  |
| **Primary Metric**           | Recall@1                        |
| **Ranking Metric**           | MRR@10                          |

### Results

| Metric        | Score  |
| -------------- | ------ |
| **Recall@1**   | 61.40% |
| **Recall@5**   | 88.30% |
| **Recall@10**  | 92.90% |
| **MRR@10**     | 72.68% |

This benchmark is intentionally more difficult because incorrect candidates can be semantically close to the correct document — the kind of near-miss confusion that matters most in production RAG pipelines.

## Benchmark Comparison

| Metric        | Easy    | Hard   |
| -------------- | ------- | ------ |
| **Recall@1**   | 96.70%  | 61.40% |
| **Recall@5**   | 99.90%  | 88.30% |
| **Recall@10**  | 100.00% | 92.90% |
| **MRR@10**     | 98.15%  | 72.68% |

The gap between the two evaluations is a useful signal of how the model behaves once results get semantically crowded, rather than an artifact of an easy test set.

## Semantic Analysis

A detailed failure analysis was run on the hard benchmark results:

| Analysis                            | Result  |
| ------------------------------------- | ------- |
| **Queries analyzed**                  | 1,000   |
| **Semantic hard negatives**           | 386     |
| **Average positive similarity**       | 0.6894  |
| **Average top-negative similarity**   | 0.6534  |
| **Average separation margin**         | 0.0360  |
| **Potential noisy positives**         | 0       |
| **Potential duplicates**              | 0       |

**Key finding:** The average similarity margin between the positive document and the strongest negative candidate was only **0.0360**. This indicates the remaining challenge is primarily **fine-grained semantic separation**, not simply filtering out unrelated content — which directly informs the hard-negative mining plan in the [Roadmap](#roadmap).

---

# Development Architecture

Silas follows an evaluation-driven development process:

```text
                 BASE MODEL
                     │
                     ▼
            ┌─────────────────┐
            │ Baseline Test   │
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Hard Benchmark  │
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Failure Analysis│
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Hard Negatives  │
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Fine-Tuning     │
            └────────┬────────┘
                     │
                     ▼
            ┌─────────────────┐
            │ Validation      │
            └────────┬────────┘
                     │
                     ▼
            PRODUCTION MODEL
```

---

# Quick Start

## Install

```bash
pip install -U sentence-transformers
```

## Load the Model

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "alfotech/silas-embedding-0.6b"
)
```

## Generate Embeddings

```python
texts = [
    "Artificial intelligence improves information retrieval.",
    "Vector embeddings represent semantic relationships."
]

embeddings = model.encode(
    texts,
    normalize_embeddings=True
)

print(embeddings.shape)
```

---

# Usage Guide

## Query vs. Document Embeddings

For retrieval-style workloads, query instructions can describe the intended task:

```text
Instruct: Given a search query, retrieve relevant passages.

Query: How do I optimize transformer inference?
```

```python
queries = [
    "How do I optimize transformer inference?"
]

documents = [
    "Quantization, batching, and KV-cache optimization can reduce inference latency.",
    "The stock market closed higher after strong earnings reports."
]

query_embeddings = model.encode(
    queries,
    prompt_name="query",
    normalize_embeddings=True
)

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

scores = model.similarity(
    query_embeddings,
    document_embeddings
)

print(scores)
```

> Use `prompt_name="query"` for the search-query side of asymmetric retrieval and leave documents un-prefixed. For symmetric tasks (e.g. clustering, deduplication), encode both sides the same way.

## Variable Embedding Size (Matryoshka)

Silas supports configurable embedding dimensions so storage and latency can be traded against representation capacity without re-encoding your corpus with a different model:

```python
embeddings = model.encode(
    ["Efficient vectors reduce storage requirements."],
    normalize_embeddings=True,
    truncate_dim=256
)
```

Supported range: **32 → 1024 dimensions**.

Use larger representations when maximizing representation capacity is important (e.g. hard semantic discrimination), or smaller vectors when storage and throughput are the priority (e.g. large-scale first-pass retrieval).

## Batch Processing

```python
embeddings = model.encode(
    large_text_list,
    batch_size=32,
    normalize_embeddings=True,
    show_progress_bar=True
)
```

For corpora in the millions of documents, encode in chunks and stream directly into your vector database's bulk-insert API rather than holding all vectors in memory at once.

---

# Production Deployment

## Production Starting Point

| Configuration           | Recommended Value |
| ------------------------- | ------------------ |
| **Sequence Length**        | 512                |
| **Batch Size**              | 32                 |
| **Normalization**           | Enabled            |
| **Similarity Metric**       | Cosine             |
| **Default Dimension**       | 1024               |

These are recommended starting points — production workloads should be benchmarked using representative data before locking in a configuration.

## Serving Options

Silas is a standard `sentence-transformers`-compatible model, so it can be served with:

- **Sentence Transformers**, directly in a Python service, for simplest integration and full control over batching.
- **Hugging Face Text Embeddings Inference (TEI)** or similar dedicated embedding servers, for higher-throughput, lower-latency serving behind a REST/gRPC endpoint.
- **ONNX / quantized export**, where CPU-only or edge deployment is required and GPU serving isn't available.

Choice of serving stack should be validated against your own latency, throughput, and hardware constraints — figures above are configuration defaults, not deployment benchmarks.

## Integration

```text
Application
     │
     ▼
Silas Embedding
     │
     ▼
Vector Database
     │
 ┌───┼───────────────┐
 ▼   ▼               ▼
FAISS Qdrant      pgvector
     │
     ▼
Nearest Neighbors
     │
     ▼
AI Application
```

---

# Intended Use

Silas is intended as the **representation layer** in semantic search, RAG, knowledge-base retrieval, and similarity-matching systems, primarily where:

- Text needs to be compared or retrieved by meaning rather than exact keyword match.
- A downstream ranking, generation, or filtering step consumes the retrieved candidates (Silas returns similarity, not a final answer).
- Multilingual input is expected, or embedding size needs to be tuned per deployment tier.

**Out of scope:** Silas does not verify factual correctness, does not perform classification or generation on its own, and should not be used as a sole safety or content-moderation filter — similarity scores reflect semantic closeness, not truth or safety.

---

# Bias, Risks & Limitations

Silas is currently evaluated primarily through embedding and semantic-matching experiments. Current limitations include:

- External leaderboard evaluation has not been independently performed.
- Long-context behavior has not been independently re-benchmarked across the entire 32K context window.
- Performance may vary across domains not represented in the benchmark corpus.
- Multilingual performance should be validated against the specific languages relevant to the target application — "100+ languages supported" reflects the base model's training, not per-language benchmarking by Alfo Tech Industries.
- Similarity scores do not represent factual correctness, and retrieved-but-similar text can still be wrong, biased, or outdated relative to the query's intent.

For high-stakes applications (legal, medical, financial, safety-critical), validate Silas using domain-specific evaluation datasets and appropriate system-level safeguards rather than relying on the benchmarks above alone.

---

# Current Development Status

| Component                          | Status     |
| ------------------------------------ | ---------- |
| Base Model Integration               | ✅ Complete |
| Baseline Evaluation                  | ✅ Complete |
| Easy Benchmark                       | ✅ Complete |
| Hard Benchmark                       | ✅ Complete |
| Failure Analysis                     | ✅ Complete |
| Semantic Hard-Negative Discovery     | ✅ Complete |
| Retrieval Fine-Tuning                | 🔄 Ongoing |
| Extended Blind Evaluation            | 🔄 Planned |
| Production Optimization              | 🔄 Planned |

---

# Roadmap

**Model Quality**
- Larger curated datasets
- Improved semantic hard-negative mining
- Stronger domain adaptation
- Broader multilingual evaluation

**Benchmarking**
- Blind evaluation sets
- Additional embedding benchmarks
- Expanded production-scale benchmarks
- Cross-domain evaluation

**Deployment**
- Higher-throughput inference
- Optimized vector dimensions
- Serving infrastructure
- Additional vector database integrations

---

# Versioning

| Version | Status  | Notes                                                   |
| ------- | ------- | -------------------------------------------------------- |
| v1      | Current | Initial public release; baseline + hard-benchmark results above |

Future releases that materially change benchmark numbers or the recommended production config will be tagged as new versions rather than silently overwriting these results.

---

# FAQ

**Which dimension should I use?**
Start at the default 1024 for best hard-case separation. Drop to a smaller `truncate_dim` (e.g. 256–384) once you've confirmed accuracy holds on your own hard-negative style data — don't shrink dimensions before benchmarking on your corpus.

**Do I need the `query` prompt for every use case?**
Only for asymmetric retrieval (short query → long document). For symmetric comparison tasks (dedup, clustering, paraphrase matching), encode both sides without the query prompt.

**Why is Recall@1 so much lower on the hard benchmark?**
Because the hard benchmark's negatives are semantically close to the correct answer by design (see [Semantic Analysis](#semantic-analysis)) — this is expected and is the metric the roadmap's hard-negative mining work targets directly.

**Is this model safe to use as a standalone fact-checker or filter?**
No — see [Intended Use](#intended-use) and [Bias, Risks & Limitations](#bias-risks--limitations).

---

# Reproducibility

Silas follows an evaluation-first development methodology. The project tracks:

- Model configuration
- Dataset processing
- Benchmark methodology
- Failure analysis
- Inference configuration
- Production recommendations

All benchmark values presented above are measured results from the current development evaluation.

---

# Citation

## Silas Embedding

```bibtex
@misc{silas_embedding_2026,
  title={Silas Embedding 0.6B},
  author={Alfo Tech Industries},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/alfotech/silas-embedding-0.6b}
}
```

## Qwen3 Embedding

```bibtex
@article{qwen3embedding,
  title={Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models},
  author={Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and others},
  journal={arXiv preprint arXiv:2506.05176},
  year={2025}
}
```

---

# License

**Apache-2.0**

Silas is based on Qwen3-Embedding-0.6B and follows the applicable licensing requirements of the upstream model.

---

# Contact & Support

- **Organization:** Alfo Tech Industries — [github.com/Alfo-Tech-Lab](https://github.com/Alfo-Tech-Lab)
- **Issues & feedback:** open an issue on the model's Hugging Face repository discussion tab, or via the GitHub organization above.

---

<p align="center">
  <b>Silas Embedding 0.6B</b><br/>
  <b>Alfo Tech Industries</b><br/>
  <i>Production-grade embedding infrastructure for modern AI systems.</i>
  <br/><br/>
  <a href="https://huggingface.co/alfotech/silas-embedding-0.6b">Hugging Face</a>
</p>