Interactive Playground — ELOQUENCE

Test multilingual AI
before you build it

The ELOQUENCE Interactive Playground lets you experiment with production-grade, bias-controlled dialogue systems across all EU languages — no setup required. From RAG-powered customer service agents to privacy-preserving voice assistants.

Open Playground  ↗ View source on GitHub →
35+
EU Languages
4
Live Pilots
50%
Bias Reduction Target
eloquence.lt.bsc.es
Playground
Ingestion
Feedback
User
What insurance options are available for freelancers in Spain?
EuroLLM-9B · RAG · Retrieved 3 docs
Based on the enterprise knowledge base: Freelancers in Spain (autónomos) have access to three main schemes — the RETA social security system, private health insurance through mutual societies, and optional unemployment coverage through the SEPE...
GPT-3.5 EuroLLM-9B Salamandra-7B Qwen2-Audio OLMo-2
What is it

One environment.
Every component you need.

The Interactive Playground is the central demonstration and testing environment of the ELOQUENCE project. It brings together three years of EU-funded research into a single accessible interface — letting researchers, developers and industry partners validate real-world AI behaviour without writing infrastructure code.

Built on BSC's high-performance computing cluster, it runs the same models validated in four industry pilots spanning healthcare call centres, customer service, smart home assistants, and bias-aware advisory systems.

🧪
Interactive Playground
Central experimentation environment
🗣️
LLM Chat
Basic · RAG · Dialogue Manager
🎙️
Audio Models
Voice input · Qwen2-Audio
📚
Vector Store
Upload docs · LanceDB retrieval
📊
Feedback Loop
Collect · filter · export
Core capabilities

Built for serious AI work

Everything you need to go from idea to validated, industry-ready conversational AI.

01
Retrieval-Augmented Generation
Upload enterprise documents and ground LLM responses in verified knowledge. Three chunking strategies — simple, recursive, and semantic — with configurable chunk size and embedder selection. No hallucinations on domain-specific queries.
02
Multilingual Model Access
Switch between GPT-3.5, EuroLLM-9B (all 24 EU official languages), Salamandra-7B (Spanish/Catalan focus), and OLMo 2 in the same interface. Select the right model for your language and use case without rebuilding your pipeline.
03
Audio & Voice Processing
Record voice input directly in the browser via Qwen2-Audio. Streaming 200ms chunk mode for real-time conversations, or full-recording batch mode for evaluation pipelines. Supports 8 languages including Spanish, German, Chinese and Japanese.
04
Private Vector Stores
Host a vector store inside your own network and connect it to the Playground API. Sensitive enterprise data never leaves your infrastructure, while still benefiting from hosted LLMs and the retrieval pipeline. GDPR-compatible by design.
05
Batch Evaluation Pipeline
Upload structured JSON conversation files and run them against any model configuration in a single pass. Download full histories with model responses and retrieved documents — ready for benchmarking and fine-tuning.
06
Structured Feedback Collection
Every interaction can receive user feedback — binary helpfulness flags plus free-text comments. Filter by model, prompt, or retrieval context and export as JSON for training data prep, model comparison, or quality tracking over time.
Real-world validation

Tested in production environments

The Interactive Playground is the operational backbone of four EU-validated pilots. Each addresses a distinct industry challenge using the same platform — the system you experiment with is identical to what our consortium partners deployed and evaluated in the field.

Healthcare
AI-Supervised Emergency Call Centres
LLM-powered supervision system for healthcare call centres supporting parents with newborn care queries. Triages urgency, collects structured information, and provides medical staff with context summaries before handoff.
150
Simulated dialogues
+20%
Target: first-call resolution
Enterprise
Self-Adapting Virtual Agents for Customer Service
RAG-powered agents trained on cross-industry datasets — finance, insurance, utilities — that generalise to new enterprises without fine-tuning. Understands user goals, makes API calls, and responds with verified, empathetic answers.
+50%
Target: customer satisfaction
+20%
Call containment target
Smart Home
Privacy-Preserving Home Voice Assistants
Federated learning framework for Telefónica's smart home platform. Models train collaboratively across devices without sharing raw user data — only model parameters reach the aggregation server. Includes keyword spotting, ASR, and speaker diarisation.
FL
Federated learning
Zero
Raw data leaves device
Bias Detection
Socially-Aware Career Counselling AI
Advisory agent for career and university recommendations that detects and mitigates cultural, gender, religious and ethnic biases. Uses prompt engineering and synthetic persona datasets to surface biases invisible in standard evaluation.
+24
EU languages tested
50%
Target: bias reduction
Supported models

European-first model stack,
open and reproducible

EuroLLM-9B
EuroLLM Team · Sept 2024 · 9B params
EU24 EU langsopen
Salamandra-7B
BSC + IULA · Feb 2025 · 2B / 7B / 40B
EU35 langsApache 2.0
OLMo 2
Allen Institute for AI · Jan 2025 · 7B / 13B
fully open
GPT-3.5-turbo
OpenAI · 16K context window
multilingual
Qwen2-Audio
Alibaba Cloud · July 2024 · voice + text
audio8 langs

How RAG works in the Playground

01
Upload your documents
PDF, DOCX, CSV, HTML, Markdown, TXT — the Ingestion tab accepts any structured enterprise document. Set chunk size, splitting strategy, and embedder.
02
Embed & index
Documents are split and embedded using MiniLM-L6, MPNet-base, or Ada-002. Vectors are stored in a LanceDB index — private to your network or shared on BSC infrastructure.
03
Query with grounding
User queries hit the vector store first. The top-k relevant passages are retrieved and injected into the LLM context window alongside the question.
04
Verified, auditable response
The LLM generates a response grounded in your documents. Retrieved passages are visible in the UI and API response, enabling full auditability.
Get started

Ready to experiment?

The Playground is live and publicly accessible. Access is credential-protected to ensure system stability — request access or self-host from source.

🔑
Request Access to the Hosted Playground
The live instance runs on BSC's HPC infrastructure with all models pre-loaded. Contact us to request credentials — access is validated and granted within 48 hours.
Request credentials →
⚙️
Self-Host from Source
Full source code, Docker images, and configuration files are publicly available. Run the Playground on your own infrastructure with your own models.
# Clone & deploy git clone github.com/ELOQUENCEAI/ eloquence-interactive-playground docker build -t eloquence-ip . docker run -d --net="host" \ -v /data:/data eloquence-ip

Full docs on GitHub →
Backed by 16 research & industry partners across Europe
Telefonica IDIAP CNR BSC FBK Brunel University London University of Essex UNS FTN EUI BUT Brno Privanova InoSens Transformation Lighthouse GrantXpert Omilia Synelixis Telefonica IDIAP CNR BSC FBK Brunel University London University of Essex UNS FTN EUI BUT Brno Privanova InoSens Transformation Lighthouse GrantXpert Omilia Synelixis