Uniro - AI Infrastructure Management

Draft1 error

Components

Drag onto the canvas

Language

Always

Always fires — use as the catch-all on a default decision.

Language detector

Fires when the prompt language matches the configured ISO code.

Topic & intent

Embedding classifier

Sentence-transformer classifier (needs sentence-transformers)

MMLU domain

Domain classifier with 57 built-in categories

KB (in-mem)

Tag message with nearest KB entry

Preference (LLM)

LLM-as-router (uses sparingly — per-request LLM call)

Complexity

Complexity classifier

Tiers prompts as easy / medium / hard

Token estimator

Estimate token count from char length

Reasoning marker (legacy)

DEPRECATED — use `structure` with a sequence rule instead.

Fact-check heuristic

Flags messages with N+ unverifiable claims

Content & modality

Structure

Feature × source rules (count/exists/density/sequence over regex/keyword-set).

Modality

text / image / audio / video / multimodal

Keyword match

Keyword-list match via bm25 / ngram / fuzzy.

Identity & role

Request metadata

Pull a field from request.metadata

AuthZ (RBAC)

Kubernetes-style RBAC from request.metadata

Conversation

Conversation depth

# of turns so far in this thread

Re-ask

Latest user message similar to a recent prior

Feedback (sentiment)

Sentiment of latest user message

Time

Time of day

Route by clock — peak vs off-hours

Safety

PII regex

Regex-based PII (email/phone/ssn/cc)

Jailbreak (regex)

Pattern-based prompt injection detection

Partition

Mutually-exclusive bucket from signal values

Score

Weighted sum

Linear combination of signal scores

Round-robin

Distribute load across N items evenly

Mapping

Threshold bands

Map a continuous score to named bands

Route

Routing rule (when → model)

Model

Add a model for dispatch

Semantic cache

Check response cache before calling the model

Response jailbreak

Screen model output for injection attempts

PII redact

Remove PII from request and/or response

Prompt cache

Cache prompt tokens (OpenAI/AWS only)

Cost cap

Enforce per-decision cost ceiling

System prompt

Inject a system prompt template

Hallucination guard

Verify factual claims in the response

Best-of-N

Sample N responses and pick the best by score

Signal Extraction

Projection Coordination

Decision Making

Model Selection

Plugin Chain

router.yaml

live preview

name: my-router
defaults:
  on_no_match: route_to_default
signals:
  - name: any
    type: always
decisions:
  - name: default
    rules:
      type: always
      name: any
    priority: 0

Decision `default` is missing a model.