Support GPU inclusion/exclusion preferences in natural language prompts

[anfredette]

Summary

Users should be able to specify which GPUs they want to include or exclude in their deployment recommendations through natural language prompts.

Current Behavior

• The DeploymentIntent schema has a single preferred_gpu_type field that supports one GPU preference (e.g., "H100" or "Any GPU")
• The capacity planner filters by this single preferred GPU type in Python after querying the database
• There is no support for:
  • Multiple GPU preferences (e.g., "H100 or H200")
  • GPU exclusions (e.g., "don't use L4")

Desired Behavior

Users should be able to express GPU preferences in natural language like:

• "I would like to use H100 or H200 GPUs."
• "I don't want to use L4 hardware."
• "Only use A100 or H100 GPUs, but not L4s."

Proposed Implementation

1. Schema Changes (backend/src/context_intent/schema.py)

Replace the single preferred_gpu_type field with two new fields in DeploymentIntent:

gpu_include: list[str] = Field(
    default_factory=list,
    description="GPU types to include (empty = all GPUs allowed)"
)
gpu_exclude: list[str] = Field(
    default_factory=list,
    description="GPU types to exclude"
)

2. LLM Prompt Changes (backend/src/llm/prompts.py)

Update the intent extraction prompt and schema to:

• Extract gpu_include list when user specifies preferred GPUs
• Extract gpu_exclude list when user specifies GPUs to avoid
• Add examples for both inclusion and exclusion patterns

3. Database Query Changes (backend/src/knowledge_base/benchmarks.py)

Modify find_configurations_meeting_slo() to accept optional GPU filter parameters:

def find_configurations_meeting_slo(
    self,
    ...,
    gpu_include: list[str] | None = None,  # Filter to only these GPUs
    gpu_exclude: list[str] | None = None,  # Exclude these GPUs
) -> list[BenchmarkData]:

Add SQL WHERE clauses to filter at the database level:

• AND hardware IN (...) when gpu_include is provided
• AND hardware NOT IN (...) when gpu_exclude is provided

4. Capacity Planner Changes (backend/src/recommendation/capacity_planner.py)

• Pass the new GPU filter parameters from DeploymentIntent to find_configurations_meeting_slo()
• Remove the post-query Python filtering for preferred_gpu_type (lines 191-205)

Benefits

• More efficient: Filtering at DB level reduces data transfer
• More flexible: Supports both inclusions and exclusions
• Better UX: Users can express complex GPU preferences naturally

Acceptance Criteria

• User can specify multiple preferred GPU types in natural language
• User can exclude specific GPU types in natural language
• GPU filtering happens at the database query level
• Estimated benchmarks (JSON) also respect GPU include/exclude filters
• Existing behavior preserved when no GPU preferences specified

[anfredette]

Also, show the hardware preferences on the Technical Specification tab.