feat(agentic): tune VLM params and enable reasoning mode

Increase AGENTIC_VLM_MAX_TOKENS default from 1024 to 16384 to accommodate
Nemotron Omni reasoning token budget. Add temperature (0.6) and top_p (0.95)
fields to AgenticVLMConfig, exposed as AGENTIC_VLM_TEMPERATURE / AGENTIC_VLM_TOP_P
env vars, per the Nemotron Omni reasoning-mode defaults in docs/vlm.md.
Prepend /think to VLM_TASK_SYSTEM_PROMPT to trigger chain-of-thought reasoning
on visual question answering tasks.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Nikhil Kulkarni <nikkulkarni@nvidia.com>

Author: nv-nikkulkarni

deploy/compose/docker-compose-rag-server.yaml

@@ -256,7 +256,9 @@ services:
       AGENTIC_VLM_SERVERURL: ${AGENTIC_VLM_SERVERURL:-${APP_VLM_SERVERURL:-""}}
       AGENTIC_VLM_MODEL: ${AGENTIC_VLM_MODEL:-${APP_VLM_MODELNAME:-""}}
       AGENTIC_VLM_APIKEY: ${AGENTIC_VLM_APIKEY:-${APP_VLM_APIKEY:-""}}
-      AGENTIC_VLM_MAX_TOKENS: ${AGENTIC_VLM_MAX_TOKENS:-1024}
+      AGENTIC_VLM_MAX_TOKENS: ${AGENTIC_VLM_MAX_TOKENS:-16384}
+      AGENTIC_VLM_TEMPERATURE: ${AGENTIC_VLM_TEMPERATURE:-0.6}
+      AGENTIC_VLM_TOP_P: ${AGENTIC_VLM_TOP_P:-0.95}
       AGENTIC_VLM_ENABLE_VISUAL_VERIFY: ${AGENTIC_VLM_ENABLE_VISUAL_VERIFY:-false}
 
     ports:

deploy/compose/nvdev.env

@@ -152,4 +152,6 @@ export AGENTIC_CONTEXT_MAX_TOKENS=100000
 # export AGENTIC_VLM_ENABLED=true
 # export AGENTIC_VLM_SERVERURL=https://integrate.api.nvidia.com/v1
 # export AGENTIC_VLM_MODEL=nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
-# export AGENTIC_VLM_MAX_TOKENS=1024
+# export AGENTIC_VLM_MAX_TOKENS=16384
+# export AGENTIC_VLM_TEMPERATURE=0.6
+# export AGENTIC_VLM_TOP_P=0.95

src/nvidia_rag/rag_server/agentic_rag/builder.py

@@ -361,17 +361,21 @@ def _make_vlm_client(vlm_cfg: Any, rag_config: Any) -> Any:
 
     api_key = vlm_cfg.get_api_key() or rag_config.llm.get_api_key()
     logger.debug(
-        "Creating agentic VLM client: model=%s, url=%s, max_tokens=%s",
+        "Creating agentic VLM client: model=%s, url=%s, max_tokens=%s, temperature=%s, top_p=%s",
         vlm_cfg.model_name,
         vlm_cfg.server_url or "(api-catalog)",
         vlm_cfg.max_tokens,
+        vlm_cfg.temperature,
+        vlm_cfg.top_p,
     )
     return get_llm(
         config=rag_config,
         model=vlm_cfg.model_name,
         llm_endpoint=vlm_cfg.server_url,
         api_key=api_key,
         max_tokens=vlm_cfg.max_tokens,
+        temperature=vlm_cfg.temperature,
+        top_p=vlm_cfg.top_p,
     )
 
 

src/nvidia_rag/rag_server/agentic_rag/prompt.py

@@ -305,7 +305,8 @@
 # directly in plain text and the caller wraps it in a completeness dict.
 # =============================================================================
 
-VLM_TASK_SYSTEM_PROMPT = """You are an expert at reading charts, diagrams, tables, and images.
+VLM_TASK_SYSTEM_PROMPT = """/think
+You are an expert at reading charts, diagrams, tables, and images.
 You are given an image and a specific question about its content.
 Answer the question directly and precisely based only on what is visually present in the image.
 Report numbers, labels, and values exactly as they appear — do not round or reformat.

src/nvidia_rag/utils/agentic_rag_config.py

@@ -128,12 +128,17 @@ class AgenticVLMConfig(_ConfigBase):
     relying on the pre-generated text descriptions produced at ingest time.
     The same VLM routing applies to verify_execute tasks.
 
+    Generation parameters follow the Nemotron Omni reasoning-mode defaults
+    documented in docs/vlm.md: temperature=0.6, top_p=0.95, max_tokens=16384.
+
     Env vars:
         AGENTIC_VLM_ENABLED              — master switch (default: false)
         AGENTIC_VLM_SERVERURL            — VLM endpoint URL
         AGENTIC_VLM_MODEL                — VLM model name
         AGENTIC_VLM_APIKEY               — optional API key override
-        AGENTIC_VLM_MAX_TOKENS           — max tokens per VLM response (default: 1024)
+        AGENTIC_VLM_MAX_TOKENS           — max tokens per VLM response (default: 16384)
+        AGENTIC_VLM_TEMPERATURE          — sampling temperature (default: 0.6)
+        AGENTIC_VLM_TOP_P                — nucleus sampling mass (default: 0.95)
         AGENTIC_VLM_ENABLE_VISUAL_VERIFY — reserved for future visual fact-checking
                                            in the verify node (default: false)
     """
@@ -154,17 +159,33 @@ class AgenticVLMConfig(_ConfigBase):
     model_name: str = Field(
         default="",
         env="AGENTIC_VLM_MODEL",
-        description="VLM model name (e.g. nvidia/llama-3.2-90b-vision-instruct).",
+        description="VLM model name (e.g. nvidia/nemotron-3-nano-omni-30b-a3b-reasoning).",
     )
     api_key: SecretStr | None = Field(
         default=None,
         env="AGENTIC_VLM_APIKEY",
         description="API key for the VLM endpoint. Falls back to NVIDIA_API_KEY if unset.",
     )
     max_tokens: int = Field(
-        default=1024,
+        default=16384,
         env="AGENTIC_VLM_MAX_TOKENS",
-        description="Max generated tokens for VLM responses.",
+        description="Max generated tokens for VLM responses (includes reasoning tokens).",
+    )
+    temperature: float = Field(
+        default=0.6,
+        env="AGENTIC_VLM_TEMPERATURE",
+        description=(
+            "Sampling temperature for VLM responses. "
+            "0.6 is the Nemotron Omni reasoning-mode default (docs/vlm.md)."
+        ),
+    )
+    top_p: float = Field(
+        default=0.95,
+        env="AGENTIC_VLM_TOP_P",
+        description=(
+            "Nucleus sampling mass for VLM responses. "
+            "0.95 is the Nemotron Omni reasoning-mode default (docs/vlm.md)."
+        ),
     )
     enable_visual_verify: bool = Field(
         default=False,