Pleias RAG 350M

🚀 Pleias-RAG-350m

Pleias-RAG-350M is a 350-million-parameter Small Reasoning Model. It's trained for retrieval-augmented general (RAG), search, and source summarization. Along with Pleias-RAG-1B, it belongs to the first generation of Pleias specialized reasoning models. This model outperforms most SLMs (4 billion parameters and below) on standardized benchmarks for retrieval-augmented general tasks and is a cost - effective alternative to popular larger models.

Full model report

🚀 Quick Start

The easiest way to deploy Pleias-RAG-350M is through our official library. It features an API-like workflow with standardized export of the structured reasoning/answer output into json format. A Colab Notebook is available for easy tests and experimentations.

✨ Features

Citation support

Pleias-RAG-350M natively generates grounded answers based on excerpts and citations extracted from the provided sources, using a custom syntax inspired by Wikipedia (). It is one of the few open weights models to date developed with this feature and the first one designed for actual deployment.

In contrast with the Anthropic approach (Citation mode), citations are integrally generated by the model and are not the product of external chunking. As a result, we can provide another desirable feature to simplify source checking: citation shortening for longer excerpts (using "(…)").

RAG reasoning

Pleias-RAG-350M generates specific reasoning sequences incorporating several proto - agentic abilities for RAG applications. The model can make a series of decisions directly:

• Assessing whether the query is understandable.
• Assessing whether the query is trivial enough to not require a lengthy pre - analysis (adjustable reasoning)
• Assessing whether the sources contain enough input to generate a grounded answer.

The structured reasoning trace includes the following steps:

• Language detection of the query. The model will always strive to answer in the language of the original query.
• Query analysis and associated query report. The analysis can lead to a standard answer, a shortening reasoning trace/answer for trivial questions, a reformulated query, or a refusal (that could, in the context of the application, be transformed into user input querying).
• Source analysis and associated source report. This step evaluates the coverage and depth of the provided sources regarding the query.
• Draft of the final answer.

Multilinguality

Pleias-RAG-350M can read and write in the main European languages: French, German, Italian, Spanish, and, to a lesser extent, Polish, Latin, and Portuguese.

To date, it is the only small language model with negligible loss of performance in leading European languages for RAG - related tasks. On a translated set of HotPotQA, we observed a significant drop of performance in most SLMs from 10% to 30 - 35% for sub - 1B models.

We expect the results of any standard English evaluation on Pleias RAG models to be largely transferable to the main European languages, limiting the costs of evaluation and deployment in multilingual settings.

📦 Installation

The installation can be done through the official library. You can access it via our official library.

💻 Usage Examples

Basic Usage

rag = RAGWithCitations("PleIAs/Pleias-RAG-350M")

# Define query and sources
query = "What is the capital of France?"
sources = [
    {
        "text": "Paris is the capital and most populous city of France. With an estimated population of 2,140,526 residents as of January 2019, Paris is the center of the Île-de-France metropolitan area and the hub of French economic, political, and cultural life. The city's landmarks, including the Eiffel Tower, Arc de Triomphe, and Cathedral of Notre-Dame, make it one of the world's most visited tourist destinations.",
        "metadata": {"source": "Geographic Encyclopedia", "reliability": "high"}
    },
    {
        "text": "The Eiffel Tower is located in Paris, France. It was constructed from 1887 to 1889 as the entrance to the 1889 World's Fair and was initially criticized by some of France's leading artists and intellectuals for its design. Standing at 324 meters (1,063 ft) tall, it was the tallest man-made structure in the world until the completion of the Chrysler Building in New York City in 1930. The tower receives about 7 million visitors annually and has become an iconic symbol of Paris and France.",
        "metadata": {"source": "Travel Guide", "year": 2020}
    }
]

# Generate a response
response = rag.generate(query, sources)

# Print the final answer with citations
print(response["processed"]["clean_answer"])

With expected output:

The capital of France is Paris. This can be confirmed by the fact that Paris is explicitly stated to be "the capital and most populous city of France" [1].

**Citations**
[1] "Paris is the capital and most populous city of France" [Source 1]

📚 Documentation

Training

Pleias-RAG-350M is trained on a large synthetic dataset emulating retrieval of a wide variety of multilingual open sources from Common Corpus. They provide native support for citation and grounding with literal quotes. Following the latest trends of agentification, the models reintegrate multiple features associated with RAG workflows such as query routing, query reformulation, source reranking.

Evaluation

Pleias-RAG-350M has been evaluated on three standard RAG benchmarks, 2wiki, HotpotQA, and MuSique.

All the benchmarks only assess the "trivial" mode on questions requiring some form of multi - hop reasoning over sources (answer disseminated into different sources) as well as discrimination of distractor sources.

Pleias-RAG-350M is not simply a cost - effective version of larger models. We found it can answer correctly to several hundred questions from HotPotQA that neither Llama-3 - 8b nor Qwen-2.5 - 7b could solve. Consequently, we encourage its use as part of multi - model RAG systems.

📄 License

The license for this project is apache - 2.0.

Model Information

Github repository: https://github.com/Pleias/Pleias-RAG-Library