Roadmap
What we're building toward. Priorities shift based on customer demand - if something here matters to you or something missing should be on the list, .
Local routing policy
The measurement pipeline is the foundation. The next major milestone turns that measurement into action: generating a routing policy from your benchmark results and shipping it as a local JSON config your agent executes. No network call at routing time. No runtime dependency on XReduce.
This is the core architectural bet - routing lives in your orchestrator, not in our service.
Serving runtime support
Most production self-hosted inference runs through a serving runtime - vLLM, TGI, Ollama, or an in-house OpenAI-compatible endpoint. Today XReduce loads models in-process; expanding to hit existing inference endpoints is one of our highest-priority engineering tracks.
See Supported models → Serving runtimes for the full matrix of what's in progress and what's planned.
Evaluation depth
A deeper evaluation surface is the other priority track alongside routing. Specific directions:
- Configurable LLM-as-judge. Choose any supported model as the judge, supply custom judging prompts, set correctness cutoffs for numeric scores.
- Additional metrics. BLEU, ROUGE, METEOR, and Ragas metrics alongside the current LLM-as-judge and task-type-based scoring.
- RAG evaluation. Faithfulness and answer relevance scoring for retrieval-augmented generation pipelines.
- Structured output scoring. Richer evaluation for schema-constrained outputs - JSON match with field-level precision, recall, and F1.
- Multi-turn conversation evaluation. Scoring logic that handles dialogue across multiple turns, not just single-shot prompts.
Model coverage expansion
See the full Supported models matrix for what's verified, compatible, in progress, and planned across distilled specialists, open-weight families, model formats, and hardware targets.
What this list isn't
This isn't a committed timeline or a feature queue with dates attached. It's the direction we're building in, ordered roughly by priority. The ordering can change if a customer asks for something loudly enough. That's part of being in early access - we're calibrating the product against the teams we're working with.
If you want specifics on when any item is likely to ship, the honest answer is a conversation, not a date on a web page. .