csml-rpi/ai-cfd-scientist/cfd-foamagent

{
  "status": "success | failed | timeout",
  "case_dir": "...",
  "case_name": "backward_facing_step",
  "case_solver": "simpleFoam",
  "loop_count": 1,
  "wall_time_s": 1234.5,
  "error_logs": []
}

[1] Parse requirement       — apply parse_system_prompt → case_info {name, domain, category, solver}
[2] RAG retrieval           — scripts/rag_query.py against three indices → tutorial_reference, allrun_reference, dir_structure
[3] Decompose to subtasks   — apply decompose_system_prompt → list of (file_name, folder_name)
[4] Mesh routing            — choose standard / custom / gmsh / copied path
[5] For each subtask:       — apply INITIAL_WRITE_SYSTEM_PROMPT → write file
[6] Allrun                  — apply command_system_prompt → write Allrun, chmod +x
[7] Run                     — bash <case>/Allrun, capture log
[8] If failure, REVIEW:     — apply REVIEWER_SYSTEM_PROMPT → review_analysis
                            — apply planner_system_prompt → rewrite plan {target_files: [...]}
                            — apply EDIT_WRITE_SYSTEM_PROMPT to each → rewrite files
                            — re-run [7]; loop max max_loop times
[9] Write run_result.json

Please transform the following user requirement into a standard case description using a structured format.
The key elements should include case name, case domain, case category, and case solver.
Note: case domain must be one of {case_domain_list}.
Note: case category must be one of {case_category_list}.
Note: case solver must be one of {case_solver_list}.

User requirement: {user_requirement}.

{
  "case_name":     "<string>",
  "case_domain":   "<one of case_domain_list>",
  "case_category": "<one of case_category_list>",
  "case_solver":   "<one of case_solver_list>"
}

# 1. similar full case content (writer + reviewer reference)
python scripts/rag_query.py \
  --db openfoam_tutorials_details \
  --query "{case_solver} {user_requirement}" \
  --top 3 \
  --output <case_dir>/.foamagent_state.tutorials.json
 
# 2. directory structure of similar cases (planner reference)
python scripts/rag_query.py \
  --db openfoam_tutorials_structure \
  --query "{case_solver} {case_domain} {case_category}" \
  --top 3 \
  --output <case_dir>/.foamagent_state.dir.json
 
# 3. Allrun scripts of similar cases
python scripts/rag_query.py \
  --db openfoam_allrun_scripts \
  --query "{case_solver} {user_requirement}" \
  --top 3 \
  --output <case_dir>/.foamagent_state.allrun.json

python Foam-Agent/init_database.py

You are an experienced Planner specializing in OpenFOAM projects.
Your task is to break down the following user requirement into a series of smaller, manageable subtasks.
For each subtask, identify the file name of the OpenFOAM input file (foamfile) and the corresponding folder name where it should be stored.
Your final output must strictly follow the JSON schema below and include no additional keys or information:
 
{
  "subtasks": [
    { "file_name": "<string>", "folder_name": "<string>" }
    // ... more subtasks
  ]
}
 
Make sure that your output is valid JSON and strictly adheres to the provided schema.
Make sure you generate all the necessary files for the user's requirements.

User Requirement: {user_requirement}
 
Reference Directory Structure (similar case): {dir_structure}
 
{dir_counts_str}
 
Make sure you generate all the necessary files for the user's requirements.
Do not include any gmsh files like .geo etc. in the subtasks.
Only include blockMesh or snappyHexMesh if the user hasn't requested for gmsh mesh or user isn't using an external uploaded custom mesh.
Please generate the output as structured JSON.

You are an expert in OpenFOAM simulation and numerical modeling.
Your task is to generate a complete and functional file named: <file_name>{file_name}</file_name> within the <folder_name>{folder_name}</folder_name> directory.
Ensure all required values are present and match with the files content already generated.
Before finalizing the output, ensure:
- All necessary fields exist (e.g., if `nu` is defined in `constant/transportProperties`, it must be used correctly in `0/U`).
- Cross-check field names between different files to avoid mismatches.
- Ensure units and dimensions are correct for all physical variables.
- Ensure case solver settings are consistent with the user's requirements. Available solvers are: {case_solver}.
Provide only the code—no explanations, comments, or additional text.

User requirement: {user_requirement}
 
Tutorial reference (similar case content):
{tutorial_reference}
 
Already-written files in this case (for cross-referencing):
{written_files_ctx}
 
Generate <file_name>{file_name}</file_name> in <folder_name>{folder_name}</folder_name>.
Output only the file body — no markdown fences, no explanations.

You are an expert in OpenFOAM. The user will provide a list of available commands.
Your task is to generate only the necessary OpenFOAM commands required to create an Allrun script for the given user case, based on the provided directory structure.
Return only the list of commands—no explanations, comments, or additional text.