Configuration Reference

July 15, 2026 · View on GitHub

The AI-Q blueprint is configured through a single YAML file that defines LLMs, tools, agents, and the workflow. The NeMo Agent Toolkit reads this file at startup and wires everything together.

Config File Structure

Every config file has four top-level sections:

general:     # Telemetry, logging, front-end settings
llms:        # LLM definitions (model, endpoint, parameters)
functions:   # Tools and agents (search tools, classifiers, research agents)
workflow:    # Top-level orchestrator configuration

Environment Variable Substitution

You can reference environment variables anywhere in the YAML using shell-style syntax:

# Required variable (fails if not set)
api_key: ${NVIDIA_API_KEY}

# Variable with a default value
checkpoint_db: ${AIQ_CHECKPOINT_DB:-./checkpoints.db}

# Nested in a URL
collection_name: ${COLLECTION_NAME:-test_collection}

The syntax ${VAR_NAME} substitutes the value of the environment variable. The syntax ${VAR_NAME:-default} provides a fallback value if the variable is not set. Environment variables are typically defined in deploy/.env or .env at the project root.

`general` Section

Controls telemetry, logging, and the application front-end.

general:
  use_uvloop: true          # Use uvloop for better async performance (web mode)
  telemetry:
    logging:
      console:
        _type: console
        level: INFO          # DEBUG, INFO, WARNING, ERROR
    tracing:
      phoenix:               # Optional: Phoenix observability
        _type: phoenix
        endpoint: http://localhost:6006/v1/traces
        project: dev
  front_end:                 # Only for web/API mode
    _type: aiq_api
    runner_class: aiq_api.plugin.AIQAPIWorker
    db_url: ${NAT_JOB_STORE_DB_URL:-sqlite+aiosqlite:///./jobs.db}
    expiry_seconds: 86400
    cors:
      allow_origin_regex: 'http://localhost(:\d+)?|http://127.0.0.1(:\d+)?'
      allow_methods: [GET, POST, DELETE, OPTIONS]
      allow_headers: ["*"]
      allow_credentials: true
      expose_headers: ["*"]

Parameter	Type	Default	Description
`use_uvloop`	`bool`	`false`	Enable uvloop for improved async I/O performance. Recommended for web mode.
`telemetry.logging.console._type`	`str`	`console`	Logging backend type.
`telemetry.logging.console.level`	`str`	`INFO`	Log level: `DEBUG`, `INFO`, `WARNING`, `ERROR`.
`telemetry.tracing`	`object`	--	Optional tracing configuration (Phoenix, OpenTelemetry).
`front_end._type`	`str`	--	Front-end type. Use `aiq_api` for the web API server. Omit for CLI mode.
`front_end.db_url`	`str`	`sqlite+aiosqlite:///./jobs.db`	Database URL for async job persistence.
`front_end.expiry_seconds`	`int`	`86400`	How long completed jobs remain in the database (seconds).
`front_end.cors`	`object`	--	CORS settings for the API server.

For aiq_api, request tag enrichment for NAT-exported spans is configured via environment variables rather than YAML fields. Refer to frontends/aiq_api/README.md and the Observability guide for:

AIQ_TRACE_USER_IDENTITY_MODE
AIQ_TRACE_USER_IDENTITY_HMAC_SECRET
AIQ_TRACE_CLIENT_ID_MODE
AIQ_TRACE_CLIENT_ID_HMAC_SECRET
AIQ_TRACE_CLIENT_IP_HEADERS

`llms` Section

Defines named LLM instances. Each entry gets a user-chosen key (for example, nemotron_super_llm) that agents reference.

llms:
  nemotron_super_llm:
    _type: nim
    model_name: nvidia/nemotron-3-super-120b-a12b
    base_url: "https://integrate.api.nvidia.com/v1"
    temperature: 0.1
    top_p: 0.3
    max_tokens: 16384
    num_retries: 5
    chat_template_kwargs:
      enable_thinking: true

Parameter	Type	Default	Description
`_type`	`str`	required	LLM provider type. Use `nim` for NVIDIA NIM endpoints, `openai` for OpenAI-compatible endpoints.
`model_name`	`str`	required	Model identifier (for example, `nvidia/nemotron-3-super-120b-a12b`, `azure/openai/gpt-4.1-mini`).
`base_url`	`str`	`None`	API endpoint URL. Should always be set explicitly for NVIDIA NIM endpoints.
`api_key`	`str`	--	API key. If omitted, uses `NVIDIA_API_KEY` from the environment.
`temperature`	`float`	`None`	Sampling temperature. Lower values produce more deterministic output. When `None`, the API uses its server-side default.
`top_p`	`float`	`None`	Nucleus sampling threshold. When `None`, the API uses its server-side default.
`max_tokens`	`int`	`300`	Maximum tokens in the response. Set higher values (for example, `16384` or `128000`) for research agents.
`num_retries`	`int`	`5`	Number of retry attempts on API failure.
`chat_template_kwargs`	`object`	--	Extra arguments passed to the chat template. Use `enable_thinking: true` to activate the model's chain-of-thought reasoning.

Common LLM Configurations

Different agents benefit from different parameter profiles:

Role	Temperature	Top-p	Max Tokens	Notes
Intent classifier	`0.5`	`0.9`	`4096`	Moderate creativity for classification
Shallow researcher	`0.1`	`0.3`	`16384`	Low temperature for factual accuracy
Deep research orchestrator	`1.0`	`1.0`	`128000`	High temperature with thinking enabled for deep reasoning
Summary LLM	`0.3`	--	`100`	Conservative, short output for document summaries

`functions` Section

Defines tools and agents. Each entry has a _type field that maps to a registered NeMo Agent Toolkit plugin. The key you assign (for example, web_search_tool) becomes the name used in tools lists.

`tavily_web_search`

Web search powered by the Tavily API.

functions:
  web_search_tool:
    _type: tavily_web_search
    max_results: 5
    max_content_length: 1000

  advanced_web_search_tool:
    _type: tavily_web_search
    max_results: 2
    advanced_search: true

  proxied_web_search_tool:
    _type: tavily_web_search
    api_base_url: https://search-proxy.example.com

Parameter	Type	Default	Description
`max_results`	`int`	`3`	Maximum number of search results to return.
`include_answer`	`str`	`"advanced"`	Whether to include a synthesized answer alongside search results. Tavily returns a direct answer in addition to individual result documents.
`api_key`	`str`	`None`	Tavily API key. Falls back to `TAVILY_API_KEY` environment variable.
`max_retries`	`int`	`3`	Number of retry attempts on search failure.
`advanced_search`	`bool`	`false`	Use Tavily's advanced search mode for deeper, more thorough results.
`max_content_length`	`int`	`None`	Truncate each result's content to this many characters. Reduces token usage.
`api_base_url`	`str`	`None`	Optional custom or proxy-compatible Tavily API base URL. A non-empty value is passed to the Tavily client; `None` uses the client's default endpoint.

`exa_web_search`

Web search powered by the Exa API via langchain-exa.

functions:
  web_search_tool:
    _type: exa_web_search
    max_results: 5
    full_text: true
    max_content_length: 10000

  deep_web_search_tool:
    _type: exa_web_search
    max_results: 5
    search_type: deep

Parameter	Type	Default	Description
`max_results`	`int`	`5`	Maximum number of search results to return.
`api_key`	`str`	`None`	Exa API key. Falls back to `EXA_API_KEY` environment variable.
`max_retries`	`int`	`3`	Number of retry attempts on search failure.
`search_type`	`str`	`"auto"`	Exa search type. See options below.
`full_text`	`bool`	`false`	Return full page text for each result. Off by default because full text is expensive in tokens; when false, results use `highlights` instead.
`highlights`	`bool`	`true`	Return highlighted snippets for each result. Highlights are token-efficient and are used as the result body when `full_text` is `false`.
`max_content_length`	`int`	`10000`	Only applied when `full_text` is `true`. Truncates each result's full page text to this many characters. Set to `None` to disable truncation.

search_type options:

auto (default) -- Let Exa pick the best strategy for the query. Balances latency and recall; a safe default for general research workloads.
fast -- Optimized for low latency. Returns results quickly at the cost of recall and semantic depth. Use for interactive UIs, high-volume calls, or when the query is narrow and keyword-like.
deep -- Optimized for thoroughness. Runs a more expensive semantic search with broader retrieval. Use for research-quality queries where completeness matters more than speed.

`nimble_web_search`

Web search powered by the Nimble Search API via langchain-nimble.

functions:
  web_search_tool:
    _type: nimble_web_search
    max_results: 5
    max_content_length: 10000

  advanced_web_search_tool:
    _type: nimble_web_search
    max_results: 5
    search_depth: deep

Parameter	Type	Default	Description
`max_results`	`int`	`5`	Maximum number of search results to return.
`api_key`	`str`	`None`	Nimble API key. Falls back to `NIMBLE_API_KEY` environment variable.
`max_retries`	`int`	`3`	Number of retry attempts on search failure.
`search_depth`	`str`	`"lite"`	Nimble search depth. See options below.
`focus`	`str`	`"general"`	Nimble focus mode. See options below.
`country`	`str`	`"US"`	ISO 3166 country code passed to Nimble (e.g. `US`, `GB`, `FR`).
`locale`	`str`	`"en"`	Language/locale passed to Nimble (e.g. `en`, `fr`, `es`).
`max_content_length`	`int \| None`	`10000`	Max characters per result's page content. Set to `None` to disable truncation.

search_depth options:

lite (default) -- Returns metadata only (title, URL, description). Fastest, lowest token cost, safe default for general lookups.
fast -- Returns rich content at low latency. Enterprise-tier only; non-enterprise accounts receive a 403 with a clear entitlement message.
deep -- Returns full page content for each result. Use for research workflows that need the body text, not just URLs.

focus options:

general (default) -- Broad web/research queries. The right choice for almost all agent use.
news -- Restricts results to news-publisher sources, ordered by recency. There is no recency threshold -- older articles still appear; it changes the source mix, not the time window. (Recency windowing is a separate Nimble time_range capability that also works with focus=general; not exposed in this initial integration.)
location, shopping, geo, social -- Domain-specific routing; set only when the tool targets that domain.

focus is a workflow-config setting, not an agent-chosen parameter -- the model only passes a query, so general research queries cannot silently switch to news. Answer generation (include_answer) is not exposed in this initial integration.

`paper_search`

Academic paper search through Google Scholar using Serper, SerpAPI, or SearchAPI. All three providers are normalized to the same agent-facing result shape. Serper is the default.

functions:
  paper_search_tool:
    _type: paper_search
    provider: serper
    max_results: 5
    serper_api_key: ${SERPER_API_KEY}

Parameter	Type	Default	Description
`provider`	`str`	`serper`	Google Scholar backend: `serper`, `serpapi`, or `searchapi`.
`max_results`	`int`	`10`	Maximum number of paper results.
`serper_api_key`	`str`	`None`	Serper key for `provider: serper`. The tool also reads `SERPER_API_KEY`.
`serpapi_api_key`	`str`	`None`	SerpAPI key for `provider: serpapi`. The tool also reads `SERPAPI_API_KEY`.
`searchapi_api_key`	`str`	`None`	SearchAPI key for `provider: searchapi`. The tool also reads `SEARCHAPI_API_KEY`.
`timeout`	`int`	`30`	Timeout in seconds for search requests.

`knowledge_retrieval`

Semantic search over ingested documents. Supports LlamaIndex (local ChromaDB), Foundational RAG (hosted NVIDIA RAG Blueprint), OpenSearch (self-hosted OpenSearch or Amazon OpenSearch Serverless), and Azure AI Search.

functions:
  # LlamaIndex backend
  knowledge_search:
    _type: knowledge_retrieval
    backend: llamaindex
    collection_name: ${COLLECTION_NAME:-test_collection}
    top_k: 5
    chroma_dir: ${AIQ_CHROMA_DIR:-/tmp/chroma_data}
    generate_summary: true
    summary_model: summary_llm
    summary_db: ${AIQ_SUMMARY_DB:-sqlite+aiosqlite:///./summaries.db}

functions:
  # Foundational RAG backend
  knowledge_search:
    _type: knowledge_retrieval
    backend: foundational_rag
    collection_name: ${COLLECTION_NAME:-test_collection}
    top_k: 5
    rag_url: ${RAG_SERVER_URL:-http://localhost:8081/v1}
    ingest_url: ${RAG_INGEST_URL:-http://localhost:8082/v1}
    timeout: 300
    # verify_ssl: false # Only set to false for self-signed certs

functions:
  # Azure AI Search backend
  knowledge_search:
    _type: knowledge_retrieval
    backend: azure_ai_search
    collection_name: ${COLLECTION_NAME:-test_collection}

This example reads AZURE_SEARCH_ENDPOINT and NVIDIA_API_KEY from the environment. AZURE_SEARCH_API_KEY is optional; when absent, the adapter uses DefaultAzureCredential.

functions:
  # OpenSearch backend
  knowledge_search:
    _type: knowledge_retrieval
    backend: opensearch
    collection_name: ${COLLECTION_NAME:-test_collection}
    top_k: 5
    opensearch_url: ${OPENSEARCH_URL:-http://localhost:9200}
    opensearch_auth_type: ${OPENSEARCH_AUTH_TYPE:-none}
    opensearch_aws_region: ${AWS_REGION:-us-east-1}
    opensearch_aws_service: ${OPENSEARCH_AWS_SERVICE:-aoss}
    opensearch_index_prefix: ${OPENSEARCH_INDEX_PREFIX:-aiq}
    opensearch_ingestion_mode: ${OPENSEARCH_INGESTION_MODE:-auto}
    embed_model: ${AIQ_EMBED_MODEL:-nvidia/llama-nemotron-embed-vl-1b-v2}

Parameter	Type	Default	Description
`backend`	`str`	`llamaindex`	Backend type: `llamaindex`, `opensearch`, `foundational_rag`, or `azure_ai_search`.
`collection_name`	`str`	`default`	Name of the document collection/index.
`top_k`	`int`	`5`	Number of results to return per query.
`generate_summary`	`bool`	`false`	Generate one-sentence summaries for ingested documents.
`summary_model`	`str`	`None`	LLM reference from `llms` section. Required when `generate_summary: true`.
`summary_db`	`str`	`sqlite+aiosqlite:///./summaries.db`	Database URL for document summaries (SQLite or PostgreSQL).
`chroma_dir`	`str`	`/tmp/chroma_data`	ChromaDB persistence directory. LlamaIndex backend only.
`rag_url`	`str`	`http://localhost:8081/v1`	RAG query server URL. Foundational RAG backend only.
`ingest_url`	`str`	`http://localhost:8082/v1`	RAG ingestion server URL. Foundational RAG backend only.
`timeout`	`int`	`120`	Request timeout in seconds. Foundational RAG backend only.
`verify_ssl`	`bool`	`true`	Verify SSL certificates. Set `false` for self-signed certs. Foundational RAG backend only.
`azure_search_endpoint`	`URL`	`AZURE_SEARCH_ENDPOINT`	Azure AI Search service endpoint. Required for Azure AI Search.
`azure_search_api_key`	`SecretStr`	`AZURE_SEARCH_API_KEY`	Optional admin API key.
`azure_search_index_prefix`	`str`	`AIQ_AZURE_SEARCH_INDEX_PREFIX` or `aiq`	Deployment-unique namespace for the shared AI-Q index.
`embed_dim`	`int`	`AIQ_EMBED_DIM` or `2048`	Embedding dimensions; must match the model and existing index schema.
`opensearch_url`	`str`	`http://localhost:9200`	OpenSearch endpoint. OpenSearch backend only.
`opensearch_auth_type`	`str`	`none`	Authentication mode: `none`, `basic`, or `sigv4`.
`opensearch_username`	`str`	`None`	Username for basic authentication. Also read from `OPENSEARCH_USERNAME`.
`opensearch_password`	`str`	`None`	Password for basic authentication. Also read from `OPENSEARCH_PASSWORD`.
`opensearch_verify_certs`	`bool`	`true`	Verify OpenSearch TLS certificates. Disable only for a trusted development cluster.
`opensearch_ca_certs`	`str`	`None`	Optional custom CA bundle path.
`opensearch_aws_region`	`str`	`us-east-1`	AWS region for SigV4 authentication.
`opensearch_aws_service`	`str`	`aoss`	SigV4 service: `aoss` for Serverless or `es` for managed OpenSearch Service.
`opensearch_index_prefix`	`str`	`aiq`	Prefix for the physical index created for each AI-Q collection.
`opensearch_embedding_dim`	`int`	`2048`	Vector dimension; must match the configured embedding model.
`opensearch_ingestion_mode`	`str`	`local`	Ingestion executor: `local`, `dask`, or `auto`. `auto` uses Dask only when a scheduler address is configured.
`opensearch_dask_scheduler_address`	`str`	`None`	Dask scheduler for distributed ingestion. Also reads `NAT_DASK_SCHEDULER_ADDRESS`.
`opensearch_dask_file_transfer`	`str`	`bytes`	Send uploads to Dask workers as `bytes` or shared filesystem `paths`.
`embed_model`	`str`	`nvidia/llama-nemotron-embed-vl-1b-v2`	Embedding model for OpenSearch ingestion and retrieval.
`embed_base_url`	`str`	`https://integrate.api.nvidia.com/v1`	OpenAI-compatible embeddings endpoint.

Refer to Knowledge Layer for backend selection and the Amazon OpenSearch Serverless guide for SigV4, IAM, and AOSS setup.

`intent_classifier`

Classifies user queries as meta (conversational) or research, and determines research depth (shallow vs. deep).

functions:
  intent_classifier:
    _type: intent_classifier
    llm: nemotron_llm_intent
    tools:
      - web_search_tool
      - paper_search_tool
    verbose: true
    llm_timeout: 90

Parameter	Type	Default	Description
`llm`	`str`	required	Reference to an LLM defined in `llms` section.
`tools`	`list[str]`	`[]`	Tool references passed to the intent prompt for tool-awareness.
`verbose`	`bool`	`false`	Enable verbose logging with trace callbacks.
`llm_timeout`	`float`	`90`	Timeout in seconds for the intent classification LLM call.

`clarifier_agent`

Interactive clarification dialog for deep research queries. Asks follow-up questions to refine scope before research begins.

functions:
  clarifier_agent:
    _type: clarifier_agent
    llm: nemotron_llm
    tools:
      - web_search_tool
    max_turns: 3
    log_response_max_chars: 2000
    verbose: true

Parameter	Type	Default	Description
`llm`	`str`	required	LLM for generating clarification questions.
`tools`	`list[str]`	`[]`	Tools available for gathering context during clarification.
`exclude_tools`	`list[str]`	`[]`	Tool names to exclude when inheriting from the data source registry.
`max_turns`	`int`	`3`	Maximum number of clarification Q&A turns before auto-completing.
`log_response_max_chars`	`int`	`2000`	Maximum characters to log from LLM responses.
`verbose`	`bool`	`false`	Enable verbose logging.

`shallow_research_agent`

Fast, single-pass research agent that produces citation-backed answers in one tool-calling loop.

functions:
  shallow_research_agent:
    _type: shallow_research_agent
    llm: nemotron_llm
    tools:
      - web_search_tool
      - knowledge_search
    max_llm_turns: 10
    max_tool_iterations: 5
    verbose: true

Parameter	Type	Default	Description
`llm`	`str`	required	LLM for research and synthesis.
`tools`	`list[str]`	`[]`	Search tools available to the agent.
`max_llm_turns`	`int`	`10`	Maximum number of LLM turns (includes both reasoning and tool-calling steps).
`max_tool_iterations`	`int`	`5`	Maximum tool-calling iterations before forcing synthesis.
`verbose`	`bool`	`false`	Enable verbose logging.

`deep_research_agent`

Multi-phase research agent with an orchestrator, optional advisory source router, planner, concurrent researcher workers, and final writer. The planner records an answer strategy plus structured ResearchQuery objects; the orchestrator batches those queries for researcher workers and delegates final synthesis to the writer.

functions:
  deep_research_agent:
    _type: deep_research_agent
    orchestrator_llm: nemotron_super_llm
    source_router_llm: nemotron_super_llm
    researcher_llm: nemotron_super_llm
    planner_llm: nemotron_super_llm
    writer_llm: nemotron_super_llm
    # tools omitted -> inherit every tool in data_source_registry
    exclude_tools:
      - web_search_tool
    enable_source_router: true
    domain_catalog_path: configs/domain_catalogs/deep_research_domain_catalog.yml
    enable_citation_verification: true
    # Optional config-function references; define these functions before enabling:
    # skills: deep_research_skills
    # sandbox: deep_research_sandbox
    max_research_concurrency: 6
    max_concurrent_source_tool_calls: 5
    max_source_tool_batch_size: 4
    verbose: true

Parameter	Type	Default	Description
`orchestrator_llm`	`str`	required	LLM for the orchestrator that coordinates the research workflow.
`source_router_llm`	`str`	`None`	LLM for the source-router sub-agent. Falls back to `orchestrator_llm` if not specified.
`researcher_llm`	`str`	`None`	LLM for the researcher sub-agent. Falls back to `orchestrator_llm` if not specified.
`planner_llm`	`str`	`None`	LLM for the planner sub-agent. Falls back to `orchestrator_llm` if not specified.
`writer_llm`	`str`	`None`	LLM for the final writer/synthesis sub-agent. Falls back to `orchestrator_llm` if not specified.
`tools`	`list[str]`	`[]`	Explicit callable tools. An empty list inherits all tool and function-group references in `data_source_registry`; a non-empty list bypasses inheritance.
`exclude_tools`	`list[str]`	`[]`	Exact runtime tool names removed after inherited or explicit tools are resolved.
`domain_catalog_path`	`str`	`None`	Optional YAML or JSON domain catalog used by the source router. Without one, AI-Q generates a general route from available mapped sources.
`enable_source_router`	`bool`	`true`	Run the advisory source-router sub-agent before planning. It recommends available mapped sources but does not restrict worker tool bindings.
`enable_citation_verification`	`bool`	`true`	Verify final citations against sources captured from configured tool results. Set `false` only when the active source formats are not compatible with verification.
`skills`	object or function ref	`None`	Inline `deep_research_skills` config or a reference to a config-only function of that type. Skill assignments are keyed by `researcher-agent` and `writer-agent`.
`sandbox`	object or function ref	`None`	Inline `deep_research_sandbox` config or a reference to a config-only function of that type. Enables the DeepAgents execution backend.
`max_research_concurrency`	`int`	`6`	Maximum `ResearchQuery` objects accepted and run concurrently by one `run_research_batch` call.
`max_concurrent_source_tool_calls`	`int`	`5`	Shared cap on concurrent source-tool calls across all researcher workers in the run.
`max_source_tool_batch_size`	`int`	`4`	Maximum concrete inputs accepted by a batch-capable source-tool wrapper in one call.
`verbose`	`bool`	`true`	Enable verbose logging.

data_sources request filtering happens after this configured tool set is resolved. It removes tools mapped to unselected registry sources but preserves configured tools with no source mapping. Router recommendations become ordered preferred_tools and fallback_tools guidance on each ResearchQuery; workers still receive the full request-filtered callable set. Refer to Tools and Sources and the config_domain_routing_and_skills.yml reference profile.

`workflow` Section

Defines the top-level orchestrator that wires together all agents.

workflow:
  _type: chat_deepresearcher_agent
  enable_escalation: true
  enable_clarifier: true
  use_async_deep_research: true
  max_history: 20
  verbose: true
  checkpoint_db: ${AIQ_CHECKPOINT_DB:-./checkpoints.db}

Parameter	Type	Default	Description
`_type`	`str`	required	Workflow type. Use `chat_deepresearcher_agent` for the full pipeline.
`enable_escalation`	`bool`	`true`	Allow the intent classifier to route queries to deep research. When `false`, all research queries use shallow research only.
`enable_clarifier`	`bool`	`true`	Run the clarifier agent before deep research to gather user requirements.
`use_async_deep_research`	`bool`	`false`	Submit deep research as an async background job (requires Dask scheduler).
`max_history`	`int`	`20`	Maximum number of messages to keep in conversation history before trimming.
`verbose`	`bool`	`false`	Enable verbose logging.
`checkpoint_db`	`str`	`./checkpoints.db`	SQLite path or PostgreSQL DSN for persistent conversation checkpoints.

Note: interactive_auth is a YAML-level field consumed by the CLI entry point (start_cli.sh / aiq-research), not a Pydantic field on ChatDeepResearcherConfig. It can be set in YAML config files but is not part of the workflow config class.

Annotated Core Pipeline Example

Below is a self-contained CLI configuration with web search, paper search, and clarification enabled. It intentionally does not combine the knowledge backends, MCP OAuth, guardrails, domain routing, or skills/sandbox examples; use the provided profiles in the next section as focused starting points for those capabilities.

# General settings
general:
  telemetry:
    logging:
      console:
        _type: console
        level: INFO                    # Set to DEBUG for troubleshooting

# LLM definitions
llms:
  intent_llm:                          # Used by intent classifier
    _type: nim
    model_name: nvidia/nemotron-3-super-120b-a12b
    base_url: "https://integrate.api.nvidia.com/v1"
    temperature: 0.5
    top_p: 0.9
    max_tokens: 4096
    num_retries: 5
    chat_template_kwargs:
      enable_thinking: true

  research_llm:                        # Used by shallow researcher + clarifier
    _type: nim
    model_name: nvidia/nemotron-3-super-120b-a12b
    base_url: "https://integrate.api.nvidia.com/v1"
    temperature: 0.1
    top_p: 0.3
    max_tokens: 16384
    num_retries: 5
    chat_template_kwargs:
      enable_thinking: true

  deep_llm:                            # Used by deep research orchestrator
    _type: nim
    model_name: nvidia/nemotron-3-super-120b-a12b
    base_url: "https://integrate.api.nvidia.com/v1"
    temperature: 1.0
    top_p: 1.0
    max_tokens: 128000
    num_retries: 5
    chat_template_kwargs:
      enable_thinking: true

# Tools and agents
functions:
  web_search_tool:                     # Standard web search
    _type: tavily_web_search
    max_results: 5
    max_content_length: 1000

  advanced_web_search_tool:            # Deep search (fewer results, more depth)
    _type: tavily_web_search
    max_results: 2
    advanced_search: true

  paper_search_tool:                   # Academic paper search
    _type: paper_search
    max_results: 5
    serper_api_key: ${SERPER_API_KEY}

  intent_classifier:                   # Classifies queries, routes depth
    _type: intent_classifier
    llm: intent_llm
    tools:
      - web_search_tool
      - paper_search_tool

  clarifier_agent:                     # Asks clarifying questions for deep research
    _type: clarifier_agent
    llm: research_llm
    tools:
      - web_search_tool
    max_turns: 3
    verbose: true

  shallow_research_agent:              # Fast single-pass research
    _type: shallow_research_agent
    llm: research_llm
    tools:
      - web_search_tool
    max_llm_turns: 10
    max_tool_iterations: 5

  deep_research_agent:                 # Multi-phase deep research
    _type: deep_research_agent
    orchestrator_llm: deep_llm
    researcher_llm: research_llm
    source_router_llm: research_llm
    writer_llm: deep_llm
    tools:
      - paper_search_tool
      - advanced_web_search_tool

# Top-level orchestrator
workflow:
  _type: chat_deepresearcher_agent
  enable_escalation: true              # Allow deep research routing
  enable_clarifier: true               # Ask clarifying questions first
  checkpoint_db: ${AIQ_CHECKPOINT_DB:-./checkpoints.db}

Provided Config Files

The repository includes eleven top-level workflow configurations. They are focused reference profiles, not cumulative layers, and no single profile enables every capability. Start from the profile closest to the deployment and merge only the additional sections you need.

File	Mode	Enabled behavior and opt-ins
`configs/config_cli_default.yml`	CLI	Chat pipeline with Tavily web search and clarification. No knowledge backend. Paper search is present only as a commented opt-in.
`configs/config_web_default_llamaindex.yml`	Web API	Default chat pipeline with LlamaIndex/ChromaDB knowledge retrieval and Tavily. Paper search is commented out.
`configs/config_web_azure_ai_search.yml`	Web API	Azure AI Search knowledge retrieval and web search
`configs/config_web_frag.yml`	Web API / Helm base	Foundational RAG plus Tavily. Requires separately deployed RAG query and ingestion services. Paper search is commented out.
`configs/config_web_opensearch.yml`	Web API	Built-in OpenSearch knowledge backend plus Tavily. Supports unauthenticated or basic self-hosted OpenSearch and SigV4 (`es` or `aoss`); infrastructure and credentials are deployment opt-ins.
`configs/config_frontier_models.yml`	Web API	LlamaIndex plus explicit per-agent tools, Nemotron researcher roles, and an OpenAI frontier model for orchestration/planning/writing. Requires `OPENAI_API_KEY`; paper search is commented out.
`configs/config_web_default_guardrails.yml`	Web API	LlamaIndex plus workflow Guardrails attachment. Shallow/deep middleware types are defined as capability examples; refer to Guardrails for the active attachment semantics.
`configs/config_web_frag_mcp_auth.yml`	Web API	Foundational RAG plus a protected per-user OAuth MCP source example. Requires a real protected MCP endpoint and shared token-store configuration; it is not a zero-config default.
`configs/config_domain_routing_and_skills.yml`	Direct deep-research workflow	Automatic domain routing, Tavily, DuckDuckGo news, Polymarket, LlamaIndex, enabled Serper paper search, built-in skills, and a Modal sandbox. Requires the corresponding service credentials and Modal setup.
`configs/config_openshell.yml`	Web API, experimental	Skills and artifact capture over one pre-provisioned named OpenShell sandbox. Intended for trusted single-operator use; per-job directories are not multi-tenant isolation.
`configs/config_mcp.yml`	Standalone MCP server	Public NIM and Tavily research over stateless submit, poll, and final-report tools with PostgreSQL-backed job state. Requires `NVIDIA_API_KEY`, `TAVILY_API_KEY`, and `AIQ_CHECKPOINT_DB`.

Swapping Models -- Change LLMs without touching agent code
Tools and Sources -- Enable and disable search tools
Knowledge Layer -- Configure document retrieval backends
Prompts -- Customize agent behavior through prompt templates