RAG Playground

Dataset-scoped assistants that behave like production

Upload data, index into embeddings, retrieve with rerank, then generate grounded answers with citations. Playground is a Beta surface: your LLM runs outside — Xalorra keeps context, versions, and traceable runs coherent.

Want to explore first?Start in StudioDocs
OpenAIKubernetesPostgreSQLDuckDBParquet
XALORRA PLAYGROUND
RAG Workflow
Dataset-scoped indexing, retrieval, rerank, and grounded generation with traces.
Index
Embeddings
Retrieve
Rerank
Citations
Traces

Stop demoing “smart chat”.
Ship assistants grounded in your actual data.

Index your dataset in minutes, not weeks.
Rerank to keep relevance high as data grows.
Citations + traces so teams can trust outputs.

Designed for grounded RAG workflows

Retrieval, rerank, and generation — without platform chaos.

Dataset-scoped context
RAG stays tied to your dataset versions, not random prompts and fragile notebooks.
Audit-friendly runs
Trace inputs/outputs with reproducible context so teams can debug and govern.
Production behavior
Same contracts you test in Playground are the ones you ship to production endpoints.
RAG Runtime (Beta)
Gateway-based RAG: your LLM runs outside. Xalorra keeps dataset scope, versions, and traceable runs coherent.
Embeddings index
Retrieve
Rerank
Citations
Traces
Build the assistant

Teams that build RAG

Pick a dataset + namespace and generate a domain assistant from grounded context.
Chunking + embeddings are versioned so runs stay reproducible.
Swap LLM providers without rewriting workflows.
Prefer to explore first? Open Playground.
ChatRetrieveTracesMODE playgroundMODEL gpt-4oMS 320Assistant Prompt + ContextSendYou are a dataset-scoped assistant.Answer with citations to retrieved chunks.Question:“What are the strongest survival signals in this dataset?”MODEL RESPONSE (GROUNDED)Top signals: sex=female, pclass=1, higher fare (proxy), and lower family_size.Citations: [chunk: ttnc_012] [chunk: ttnc_044] [chunk: ttnc_107]
Improve relevance

Teams that rerank retrieval

Rerank improves precision when top-k retrieval gets noisy.
Keep citations consistent even as the dataset grows.
Measure quality by looking at traces and retrieved chunks.
Prefer to explore first? Open Playground.
ChatRetrieveTracesSTAGE retrieveTOPK 8RERANK onMS 48Retrieve + Rerank (dataset-scoped)Rundataset: default.titanic_debug_1000_clean @v2query: "survival signals"retrieve: topk=8rerank: on (cross-encoder)RETRIEVED CHUNKS (RERANKED)#1ttnc_044score 0.92female passengers in higher classes...#2ttnc_107score 0.89pclass and fare correlate with...#3ttnc_012score 0.86family_size and is_alone patterns...
Govern what you ship

Teams that govern runs

Audit inputs/outputs and track dataset versions across runs.
Tenant-aware isolation to prevent cross-org leakage.
Operational clarity: what changed, when, and why.
Prefer to explore first? Open Playground.
ChatRetrieveTracesTRACES enabledCITATIONS onRUN okMS 96Traces + CitationsInspectrun_id: rag_2026_01_19_0007dataset_scope: default.titanic_debug_1000_clean @v2citations: enabledtrace_level: fullTRACE SUMMARYretrieve8 chunks (reranked)generateanswer + citationstokensprompt 1,240 • completion 286latency96 ms (pipeline)

RAG you can audit later

Grounded answers are a workflow, not a prompt.

If you care about tenant isolation, versioned datasets, and traceable runs, you want a playground that behaves like production.