Open-source MLOps Engineering Platform for retrieval + agent systems.
API-first orchestration for retrieval, training, evals, tracing, and model routing, with MCP built in for agent workflows.
Synthetic Data Lab, dual training studios, semantic cache controls, and guardrailed indexing in one operational surface.
Live search + chat are backed by the Epstein Files 1 corpus. Add ?mock=1 for offline demo mode.
A full MLOps engineering platform for retrieval + agent systems. The product center is API-first orchestration for indexing, retrieval, evals, training, tracing, and routing.
MCP support is first-class, but it sits on top of the API contract rather than replacing it. The panels below map core workflows to their matching docs destinations.
ragweld is API-first in production: routing, retrieval, evals, and observability all anchor to /api. MCP is built in for agent tooling, but the API surface is the primary integration layer.
Embedding mismatch detection, index contract locking, and guided reindex actions keep high-option indexing from drifting silently in production.
Dual training studios support live telemetry, promotion gates, rollback paths, and run inspection so quality moves are operational, not guesswork.
Graph explorer and retrieval controls are in the same workbench, so structural signals are measurable, debuggable, and part of normal iteration.
Smart gating and intensity controls decide when conversational memory is injected, while semantic cache controls token spend in repeated query patterns.
Infrastructure controls, service states, and observability integration stay in the same operator UI so debug loops stay short when routing, indexing, or model behavior shifts.
Current production surfaces across API routing, indexing, training, graph inspection, and recall policy.
Route model traffic per channel with API-first defaults, MCP overrides, and provider readiness visible in one place.
Embedding mismatch detection, index contract lock controls, and guided reindex actions reduce high-cost indexing errors.
Track completed and failed runs side-by-side with run status, duration, active model path, and promotion controls.
Control how conversational memory is indexed and injected with intensity, recency, and skip-behavior gates.
Inspect neighborhood structure directly to validate entity relationships and graph-retrieval behavior.
Recompose the workspace quickly by docking the exact tabs and diagnostics needed for the current investigation.
Three-leg retrieval is the spine: vector + sparse + graph signals fused and reranked — then surfaced through a workbench that lets you measure and iterate.
Index once, then retrieve through vector, sparse, and graph legs in parallel. Fusion and reranking happen with inspectable knobs so changes are measurable instead of opaque.
The workbench is built for repeated cycles: synthetic data generation, eval runs, tracing, and routing updates all feed the same decision surface. API routes stay primary; MCP rides on top for agent clients.
ragweld is open source, self-hostable, and built as an API-first MLOps Engineering Platform.
MIT licensed. Run it on your infrastructure. Keep your corpora, embeddings, and model traffic where you want them.
For teams pushing RAG into production: deployment patterns, hardened ops, and integration work tailored to your stack.
Explore the live demo, start from the API docs, then layer MCP integrations where they fit your agent stack.