Skip to content

Feature Model Coverage — Built-in Tools & Safeguards

This page cross-references every model name used by the Built-in Tools and Context Management & Safeguards features against the Foundry alias registry, to determine what works, what needs a new alias, and what cannot be fixed with aliases alone.

Sources: - Feature model refs: Built-in Tools Reference · Context Management & Safeguards - Foundry aliases: Model Aliases & Testing - App config overrides: mod-ask-assistant-service/helm/environments/values-mod-auh1-dev-ask.yaml


Verified live 2026-06-27 — Fixes applied 2026-06-27

Config source: kubectl exec -n ask deploy/assistant-service-intelligence -- cat /app/config/mod-auh1-dev-ask.toml Foundry model list: kubectl exec -n ask deploy/assistant-service-intelligence -- python3/v1/models No model env vars are set — all overrides come from the TOML file, not environment variables.

Patches applied:

  • Fix 1small_llm_model updated to Qwen/Qwen3-4B-Instruct-2507 in values-mod-auh1-dev-ask.yaml
  • Fix 2[mod-auh1-dev-ask.history_summary] block added to values-mod-auh1-dev-ask.yaml
  • Fix 3gpt-oss-safeguard-120b alias added to 07-foundry-litellm/01-values.yaml (also added qwen/qwen3-5-4b + qwen3-5-4b aliases) — in config but pending ArgoCD sync of Foundry (not yet live as of 2026-06-28)
  • Fix 4 — OpenAI Safety Filters: architectural decision pending (disable or replace with LLaMA Guard)
  • 📋 Fix 5 — LLaMA Guard: backlog (requires new vLLM deployment)

Live Config vs Foundry — Gap Summary

Status Feature Deployed TOML value Foundry has it?
✅ Works Most tools (chat, code, docx, data-viz, data-analysis, embedding, reranker) qwen/qwen3-5-122b-a10b, qwen/qwen3-5-122b-a10b, qwen/qwen3-embedding-0-6b, nvidia/llama-nemotron-rerank-1b-v2 ✅ all present
✅ Fixed Small model path (summarization tool) Qwen/Qwen3-4B-Instruct-2507 (was qwen/qwen3-5-4b — removed) ✅ served directly
✅ Fixed Conversation Summarizer [history_summary] model = "Qwen/Qwen3-4B-Instruct-2507" (was falling to Bedrock) ✅ served directly
⚠️ Verify Custom Policies + Prompt Injection guardrails qwen3-5-122b-a10b / qwen-3-4b-it-2507 (commit 32aaae8) bare names 400 on Foundry — verify
❌ Fails OpenAI Safety Filters omni-moderation-latest ❌ external OpenAI API — disable in air-gap
❌ Not deployed LLAMA Guard meta-llama/llama-guard-3-8b ❌ no vLLM instance — backlog

Applied Fixes (2026-06-27)

Fix 1 + Fix 2 — Patched in mod-ask-assistant-service/helm/environments/values-mod-auh1-dev-ask.yaml

# Fix 1: small_llm_model (qwen3-5-4b was removed from Foundry)
[mod-auh1-dev-ask.foundry]
small_llm_model = "Qwen/Qwen3-4B-Instruct-2507"

# Fix 2: Conversation Summarizer (no TOML block → fell to anthropic/claude-haiku-4-5 on Bedrock)
[mod-auh1-dev-ask.history_summary]
model = "Qwen/Qwen3-4B-Instruct-2507"

Fix 3 — Patched in mlops-mod-production/07-foundry-litellm/01-values.yaml

router_settings:
  model_group_alias:
    # qwen3-5-4b was removed from Foundry vLLM; redirect to the served 4B model
    qwen/qwen3-5-4b: Qwen/Qwen3-4B-Instruct-2507
    qwen3-5-4b: Qwen/Qwen3-4B-Instruct-2507
    # safeguard alias — custom policies + prompt injection guardrail (L3 hardcoded in assistant-service)
    gpt-oss-safeguard-120b: qwen/qwen3-5-122b-a10b

Fix 4 — OpenAI Safety Filters (pending architectural decision)

omni-moderation-latest calls api.openai.com — unreachable in air-gap. Options:

  1. Disable the openai_moderation guardrail in assistant config (recommended for v1)
  2. Replace with Custom Policy guardrail using gpt-oss-safeguard-120b (now available via alias)
  3. Deploy LLaMA Guard and use as the safety filter backend (see Fix 5)

Fix 5 — LLAMA Guard (backlog)

Requires a new vLLM deployment for meta-llama/Llama-Guard-3-8B in the Foundry namespace. No quick fix — log as a backlog item.


Quick Legend

Symbol Meaning
✅ Works Model served locally OR L1 config already overrides the default
⚠️ Needs alias Model name is called but not registered in Foundry — add model_group_alias
🌐 External API Routes to an internet API (OpenAI, Bedrock) — air-gap must disable or replace
❌ Not deployable No local model equivalent exists; feature must be disabled

Built-in Tools — Model Coverage

All tools that use llm_model_name, small_llm_model_name, code_generator_model_name, docx_prepare_context_model_name, or data_visualization_llm are already handled by L1 config overrides in the deployed values file — the upstream defaults (azure/gpt-5-chat, anthropic/claude-3-7-sonnet) are never actually called.

Tool / Feature Settings Field Upstream Default L1 Override (deployed) Foundry Served? Status Action needed
Web Search, Email (Gmail/Outlook), Calendar, People, Knowledge Extraction, Summarization, Translation, Database Navigator, Code Generator llm_model_name / small_llm_model_name / code_generator_model_name azure/gpt-5-chat llm_modelqwen/qwen3-5-122b-a10b
small_llm_modelQwen/Qwen3-4B-Instruct-2507
code_generator_modelqwen/qwen3-5-122b-a10b
✅ yes (all three) ✅ Works None — L1 override covers it. Default never called.
DOCX (prepare context phase) docx_prepare_context_model_name qwen/qwen3-5-122b-a10b qwen/qwen3-5-122b-a10b (same as default) ✅ yes — alias gpt-oss-120b → qwen/qwen3-5-122b-a10b ✅ Works None.
Data Visualization data_visualization_llm anthropic/claude-3-7-sonnet qwen/qwen3-5-122b-a10b ✅ yes ✅ Works None — L1 override replaces Claude with local Qwen model.
Knowledge Base & Tool Search (embedding) foundry_tool_embedding_model_name BAAI/bge-large-en-v1.5 qwen/qwen3-embedding-0-6b ✅ yes ✅ Works None — L1 override replaces BAAI embedding with local Qwen embedding.
Image Analysis app_config.image_analysis.model Qwen/Qwen3-VL-8B-Instruct Disabled by default (image_analysis=@null) ❌ no — not registered ⚠️ Needs alias if enabled If you enable image_analysis in env toml, add alias:
Qwen/Qwen3-VL-8B-Instruct: qwen/qwen3-5-122b-a10b
(Qwen3-5-122B-A10B is a VLM and handles vision tasks)
Meta Agent app_config.tooling.meta_agent.llm_model_name anthropic/claude-sonnet-4-6 Disabled by default (meta_agent=@null) ❌ no — not registered ⚠️ Needs alias if enabled If you enable meta_agent, add alias:
anthropic/claude-sonnet-4-6: qwen/qwen3-5-122b-a10b
OR set the model name directly in the env toml to a served model

Context Management — Model Coverage

Feature Config Field Model Called Foundry Served? Status Action needed
Conversation Summarizer history_summary.enabled Qwen/Qwen3-4B-Instruct-2507 (L1 override applied 2026-06-27) ✅ confirmed live in Foundry ✅ Fixed Added [mod-auh1-dev-ask.history_summary] model = "Qwen/Qwen3-4B-Instruct-2507" to values file. Previously fell to upstream default anthropic/claude-haiku-4-5 (Bedrock, not deployed).
Context Compaction Auto (75% threshold) None — local trimming logic, no LLM call N/A ✅ Works None.
Tool Output Storage tool_output_offloading.enabled None — Daytona volume N/A ✅ Works None — requires Daytona running, not a model dependency.

Safeguards — Model Coverage

Feature Model Called Foundry Served? Status Action needed
Keyword Filters (Ban List / Fuzzy Ban List) None — Trie-based, no LLM N/A ✅ Works None.
Custom Policies (Policy Guardrail) gpt-oss-safeguard-120b ⏳ alias in config (2026-06-27) → qwen/qwen3-5-122b-a10bnot yet live (Foundry ArgoCD sync pending) ⏳ Pending sync Alias added to 07-foundry-litellm/01-values.yaml. Confirmed absent from live /v1/models 2026-06-28 — will take effect after ArgoCD syncs the Foundry Application.
Prompt Injection Detection gpt-oss-safeguard-120b ⏳ same alias covers both — pending Foundry sync ⏳ Pending sync Same alias as Custom Policies — one alias fixes both guardrails once Foundry ArgoCD sync completes.
Safety Filters — Hate Speech, Sexual, Violence
(OpenAI Moderation path)
omni-moderation-latest ❌ external OpenAI API call — not routed through Foundry ❌ Air-gap incompatible OpenAI Moderation API requires internet access to api.openai.com. In MOD SC air-gap:
• Disable the openai_moderation guardrail
• Replace with llama_guard (if local LLaMA Guard is deployed) or custom policy guardrail using gpt-oss-120b
Safety Filters — LLAMA Guard path
(alternate to OpenAI Moderation)
meta-llama/llama-guard-3-8b ❌ NOT in Foundry vLLM deployment ❌ Not deployed LLaMA Guard 3-8B is not currently serving in the Foundry model_list. To use:
• Deploy a vLLM instance for meta-llama/Llama-Guard-3-8B in the Foundry namespace
• Register it and add alias meta-llama/llama-guard-3-8b: meta-llama/llama-guard-3-8b
Until then: use custom policy guardrail as the only safety filter option

Action Summary — Applied + Remaining

Priority Item Status Where
Applied qwen/qwen3-5-4b + qwen3-5-4b aliases → Qwen/Qwen3-4B-Instruct-2507 Done 2026-06-27 07-foundry-litellm/01-values.yaml
Applied small_llm_modelQwen/Qwen3-4B-Instruct-2507 Done 2026-06-27 values-mod-auh1-dev-ask.yaml
Applied [history_summary] model = "Qwen/Qwen3-4B-Instruct-2507" Done 2026-06-27 values-mod-auh1-dev-ask.yaml
Pending sync gpt-oss-safeguard-120b: qwen/qwen3-5-122b-a10b alias (+ compat redirects for qwen3-5-4b) In config 2026-06-27 — ArgoCD sync of Foundry required 07-foundry-litellm/01-values.yaml
Pending Disable openai_moderation guardrail (external API, air-gap incompatible) Architecture decision assistant-service config
📋 Backlog Deploy LLaMA Guard 3-8B vLLM + register alias Phase 2 Foundry namespace
🟡 If enabling Qwen/Qwen3-VL-8B-Instruct: qwen/qwen3-5-122b-a10b alias When Image Analysis enabled 07-foundry-litellm/01-values.yaml
🟡 If enabling anthropic/claude-sonnet-4-6: qwen/qwen3-5-122b-a10b alias When Meta Agent enabled 07-foundry-litellm/01-values.yaml

ArgoCD sync required

The Foundry alias changes (07-foundry-litellm/01-values.yaml) require an ArgoCD sync of the Foundry Application to take effect. The assistant-service TOML changes (values-mod-auh1-dev-ask.yaml) require an ArgoCD sync of the assistant-service Application (triggers Reloader rolling restart).


Complete Model Inventory (both pages)

Model name called Source feature Available in Foundry Resolution
qwen/qwen3-5-122b-a10b Main orchestrator (L1 override) ✅ registered Works
Qwen/Qwen3-4B-Instruct-2507 Small model / summaries + Conversation Summarizer (L1 override, fixed 2026-06-27) ✅ registered Works
qwen/qwen3-5-122b-a10b Code-gen, DOCX, data-analysis (L1 override) ✅ registered + alias Works
qwen/qwen3-embedding-0-6b Embeddings (L1 override) ✅ registered + alias Works
azure/gpt-5-chat Upstream default (never called — L1 overrides) ❌ not registered N/A — default shadowed
anthropic/claude-3-7-sonnet Data viz default (never called — L1 override) ❌ not registered N/A — default shadowed
BAAI/bge-large-en-v1.5 Embedding default (never called — L1 override) ❌ not registered N/A — default shadowed
anthropic/claude-haiku-4-5 Conversation Summarizer (upstream default — no longer called; overridden by [history_summary] block) N/A — shadowed by L1 override Fixed 2026-06-27
gpt-oss-safeguard-120b Custom Policies, Prompt Injection ⏳ alias in config → qwen/qwen3-5-122b-a10b (pending Foundry ArgoCD sync) Fix 3 pending — not yet live 2026-06-28
omni-moderation-latest OpenAI Safety Filters ❌ external API Disable in air-gap
meta-llama/llama-guard-3-8b LLAMA Guard Safety Filters ❌ not deployed Deploy vLLM or disable
Qwen/Qwen3-VL-8B-Instruct Image Analysis (disabled by default) ❌ not registered Alias if enabling
anthropic/claude-sonnet-4-6 Meta Agent (disabled by default) ❌ Bedrock route Alias if enabling