Venice Text-Model Routing

This skill encodes the decision logic for "which Venice text model do I call?" — the routing layer that sits above venice-chat (the call surface) and consumes venice-models (the discovery API).

Primary use case: a local agent receives a prompt, decides whether the local model can handle it, and — if not — picks the cheapest Venice model that satisfies the privacy / modality / capability requirements.

Snapshot freshness

Before applying the matrix below, read snapshots/text-routing.json. If the file is missing, or snapshot_date is older than 30 days, run python scripts/refresh_routing.py to regenerate it from GET /models?type=text + GET /models/traits?type=text. Otherwise trust the cached file — do not hit /models on every routing decision.

refresh_routing.py requires VENICE_API_KEY in the environment and rewrites both snapshots/text-routing.json and routing-matrix.md. Run it once on first install, then ~monthly (CI nightly is also fine).

When to load this skill

• Picking a Venice text model from a prompt at runtime.
• Building a local-first agent that escalates to Venice for hard prompts.
• Deciding privacy tier (anonymized vs private vs TEE vs E2EE).
• Choosing between a trait shortcut (default_reasoning, most_intelligent, …) and a hand-filtered candidate.

For the chat call surface itself, see venice-chat. For raw discovery of every model field, see venice-models.

Privacy tier ladder

Pick the least restrictive tier that satisfies the request — restricting tier shrinks the candidate pool and often raises cost.

Sources: docs.venice.ai/overview/privacy, docs.venice.ai/guides/features/tee-e2ee-models.

E2EE is not supported on /responses — route encrypted requests to /chat/completions.

Capability filters

Map prompt requirement → model_spec.capabilities flag (full list in venice-models).

Cost tiers

The Venice text catalog clusters by per-1M-token price. Buckets below match the size labels documented at docs.venice.ai/models/text. Pick the smallest bucket that hosts a model satisfying your capability filters.

Authoritative per-model pricing lives on model_spec.pricing in snapshots/text-routing.json — never hard-code dollar figures from this prose.

Routing decision tree

Walk top-down. Stop at the first rule that applies.

1. Local-first check
   - Prompt is ≤ ~500 tokens, no special-capability requirement,
     no privacy escalation, no tool calls expected
     → handle on the local model. Do not call Venice.

2. Privacy gate
   - User flagged "private" OR prompt contains regulated data (PHI, secrets, legal):
       require model_spec.privacy === "private"
       OR model id startsWith "tee-" / "e2ee-".
   - User flagged "E2EE" / "must be encrypted in transit":
       require model id startsWith "e2ee-". Skip /responses.
   - User flagged "TEE" / "verifiable inference":
       require model id startsWith "tee-".
   - Otherwise: any tier acceptable (still prefer privacy: "private" when otherwise tied).

3. Modality gate
   - Image input present  → require supportsVision (+ supportsMultipleImages if > 1).
   - Audio input present  → require supportsAudioInput.
   - Video input present  → require supportsVideoInput.

4. Capability gate
   - Tool calls expected               → require supportsFunctionCalling.
   - Code-heavy task                   → prefer optimizedForCode (or call /models?type=code).
   - Chain-of-thought / planning / hard math
                                       → require supportsReasoning;
                                         prefer supportsReasoningEffort to dial reasoning.effort.
   - Structured JSON output required   → require supportsResponseSchema.
   - Web search needed                 → require supportsWebSearch
                                         (or supportsXSearch for X/Twitter content).

5. Context size gate
   - Estimated (prompt + expected output) > availableContextTokens of the candidate
     → bump up to a model with sufficient context. Prefer ones with cache_input pricing.

6. Frontier override
   - User asked for "best", "frontier", "most intelligent", "smartest"
     → resolve trait `most_intelligent` from the snapshot. Skip cost-min step.

7. Cost minimization
   - From surviving candidates, pick the smallest cost tier (XS → S → M → L → frontier).
     Tie-break by lower output $/1M, then lower input $/1M.

8. Sanity filters (apply throughout)
   - Drop model_spec.beta === true unless your key has beta access.
   - Drop model_spec.offline === true.
   - Drop candidates whose model_spec.regionRestrictions exclude the caller.

Trait shortcuts

When the prompt maps cleanly to a named trait, skip the matrix and resolve the trait from the snapshot's traits block (sourced from GET /models/traits?type=text):

Cache the resolved trait → ID map at session start (one HTTP call) and reuse.

Local-first pattern

For a local agent driving Venice as an "escalation backend":

1. Score the prompt cheaply (locally):
   • Token count of prompt + expected output.
   • Modality signals (any image / audio / video parts).
   • Keyword signals: "code", "reason", "step by step", "private", "secret", "encrypt", "best", "frontier".
   • User-provided overrides (e.g. --model frontier, --privacy e2ee).
2. Decide local vs Venice:
   • If estimated tokens ≤ local model's comfort window AND no capability or privacy escalation → stay local.
   • Otherwise → continue to step 3.
3. Run the decision tree above to pick a Venice model.
4. Call via venice-chat:

   curl https://api.venice.ai/api/v1/chat/completions \
     -H "Authorization: Bearer $VENICE_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "<chosen id>",
       "messages": [...],
       "venice_parameters": {"include_venice_system_prompt": false}
     }'
   

5. Cap blast radius: log the chosen model + estimated cost before sending; refuse to escalate beyond the user's --max-cost ceiling.

Examples

Example 1 — local-first wins

Prompt: "Summarize this email in one sentence: …" (~200 tokens)

• Step 1 → local handles it. Do not call Venice.

Example 2 — privacy + reasoning

Prompt: "Here's our customer churn dataset (PII). Reason about which factors drive churn."

• Step 2 → privacy gate requires privacy: "private" or tee- / e2ee- prefix.
• Step 4 → reasoning required → supportsReasoning: true.
• Step 7 → among private/TEE candidates with reasoning, pick the cheapest.
• → typically a tee-* reasoning model, or traits.default_reasoning filtered to privacy: "private".

Example 3 — multi-image vision

Prompt: 3 product photos + "Compare these for build quality."

• Step 3 → supportsVision && supportsMultipleImages && maxImages >= 3.
• Step 7 → cheapest survivor → typically traits.default_vision.

Example 4 — frontier intelligence on a long doc

Prompt: "Give me your absolute best take on this 50-page legal brief."

• Step 5 → context size gate triggers (> 100K tokens).
• Step 6 → frontier override fires (most_intelligent).
• → resolve traits.most_intelligent (a Claude Opus / GPT-5.x Pro / GLM 5.1 family ID, depending on the snapshot).

Example 5 — code agent with tools

Prompt: "Refactor this repo. You have shell + edit tools."

• Step 4 → supportsFunctionCalling && optimizedForCode.
• Step 7 → cheapest survivor → traits.default_code if it also supports tools, else traits.function_calling_default filtered by optimizedForCode.

Future extensions

A scripts/route.py CLI may be added later for runtimes that prefer structured output (route.py --prompt '...' --max-cost 0.001 --need-vision → {"model_id": "...", "estimated_cost": ..., "tier": "..."}). The prose decision tree above remains the source of truth.

Sibling routing skills (venice-image-routing, venice-audio-routing, venice-video-routing) can mirror this layout when needed.

Gotchas

• Don't hard-code model IDs. Venice rolls models monthly — resolve via traits or filters against the snapshot.
• Stale snapshot lies silently. The 30-day rule is the floor; refresh sooner if you hit a 404 on a model ID.
• Privacy ≠ uncensored. most_uncensored and privacy: "private" are independent axes.
• enable_e2ee defaults to true on E2EE-capable models when the right headers are present — see venice-chat. The routing decision selects the model; the chat skill drives the handshake.
• Beta / offline models in the snapshot. Filter out model_spec.beta === true (unless your key has beta access) and model_spec.offline === true before scoring.
• Region restrictions. model_spec.regionRestrictions[] returns 403 outside the listed countries — drop those candidates if your caller is outside.
• Trait keys differ by type. Always pass ?type=text. Don't reuse image traits.

See also

• venice-models — /models, /models/traits, /models/compatibility_mapping (the discovery API this skill consumes).
• venice-chat — /chat/completions (the call surface this skill picks a model for).
• venice-auth — Bearer vs x402 wallet auth.
• venice-billing — confirming actual spend matched the routing estimate.
• venice-errors — 402 / 422 / 429 handling on routed requests.
• Per-model snapshot: snapshots/text-routing.json.
• Per-model human-readable matrix: routing-matrix.md.