testarcbuilder / services /geminiService.ts
wuhp's picture
Update services/geminiService.ts
650c136 verified
import { GoogleGenAI } from "@google/genai";
import { Node, Edge } from 'reactflow';
import { NodeData, LayerType } from '../types';
import { LAYER_DEFINITIONS } from '../constants';
// Key Management
const getEnvKey = () => {
// In Development: Returns key from .env file if available
// In Production (Docker): Returns undefined/empty because Dockerfile does not inject the ARG
// @ts-ignore
if (typeof import.meta !== 'undefined' && import.meta.env && import.meta.env.VITE_GEMINI_API_KEY) {
// @ts-ignore
return import.meta.env.VITE_GEMINI_API_KEY;
}
return '';
};
let userApiKey = typeof window !== 'undefined' ? localStorage.getItem('gemini_api_key') || '' : '';
const defaultEnvKey = getEnvKey();
export const setUserApiKey = (key: string) => {
userApiKey = key;
if (typeof window !== 'undefined') {
localStorage.setItem('gemini_api_key', key);
}
};
export const getUserApiKey = () => userApiKey || defaultEnvKey;
const getAiClient = () => {
const key = getUserApiKey();
if (!key) throw new Error("API Key is missing. Please add your Gemini API Key.");
return new GoogleGenAI({ apiKey: key });
};
const MODEL_NAME = 'gemini-2.5-flash';
export type AgentStatus = 'idle' | 'architect' | 'critic' | 'refiner' | 'debugger' | 'patcher' | 'complete' | 'error';
/**
* Internal helper to build the raw specification string.
* This ensures all data is captured before sending to the AI for refinement.
*/
const buildRawPrompt = (nodes: Node<NodeData>[], edges: Edge[]): string => {
// Sort nodes by vertical position to list them in a logical flow
const sortedNodes = [...nodes].sort((a, b) => a.position.y - b.position.y);
let rawSpec = "### Raw Architecture Data\n\n";
rawSpec += "**1. Nodes (Layers):**\n";
sortedNodes.forEach(node => {
// Format parameters nicely, handling objects
const params = Object.entries(node.data.params)
.map(([k, v]) => {
if (typeof v === 'object' && v !== null) {
return `${k}=${JSON.stringify(v)}`;
}
// Don't clutter standard output with huge code blocks, handle separately
if (k === 'definition_code' || k === 'imports') return null;
return `${k}=${v}`;
})
.filter(p => p !== null)
.join(', ');
rawSpec += `- [ID: ${node.id}] TYPE: ${node.data.type} | LABEL: ${node.data.label}\n`;
if (params) {
rawSpec += ` PARAMS: ${params}\n`;
}
// Specific instruction for Custom Layers
if (node.data.type === LayerType.CUSTOM) {
rawSpec += ` CUSTOM_NOTE: Instantiate using class '${node.data.params.class_name}' with args '${node.data.params.args}'.\n`;
if (node.data.params.imports) {
rawSpec += ` CUSTOM_IMPORTS: ${node.data.params.imports}\n`;
}
if (node.data.params.definition_code) {
rawSpec += ` CUSTOM_CODE_DEFINITION:\n${node.data.params.definition_code}\n`;
}
}
});
rawSpec += "\n**2. Connectivity (Edges):**\n";
if (edges.length === 0) {
rawSpec += "- No connections defined.\n";
} else {
edges.forEach(edge => {
const sourceNode = nodes.find(n => n.id === edge.source);
const targetNode = nodes.find(n => n.id === edge.target);
const sourceName = sourceNode ? `${sourceNode.data.label} (ID:${sourceNode.id})` : edge.source;
const targetName = targetNode ? `${targetNode.data.label} (ID:${targetNode.id})` : edge.target;
rawSpec += `- ${sourceName} -> ${targetName}\n`;
});
}
return rawSpec;
};
/**
* Generates a polished, professional prompt using the AI.
* It takes the raw hardcoded spec and asks the AI to format it perfectly for a coding LLM.
*/
export const generateRefinedPrompt = async (nodes: Node<NodeData>[], edges: Edge[]): Promise<string> => {
const ai = getAiClient();
const rawSpec = buildRawPrompt(nodes, edges);
const systemPrompt = `
You are an expert AI Prompt Engineer for Deep Learning.
Your goal is to take a raw, technical neural network specification and rewrite it into a
perfect, professional, and detailed prompt that another AI (like a coding assistant) could use to write flawless PyTorch code.
Input Raw Specification:
${rawSpec}
Instructions:
1. Start the output with: "You are an expert Deep Learning Engineer. Please write a complete, runnable PyTorch model code for the following neural network architecture:"
2. Create a section "Architecture Specification".
3. List "1. Layers (Nodes)" cleanly. Include ID, Type, and Parameters.
4. List "2. Connectivity (Forward Pass Flow)" cleanly.
- Explicitly describe merge points (e.g. "Node X receives inputs from A and B. Handle this merge...").
- Note specific handling for complex layers like CrossAttention (needs Query + Key/Value) or SAM Decoders.
5. Create a section "Implementation Requirements" with standard PyTorch best practices (nn.Module, forward method, correct shapes).
6. If CUSTOM_CODE_DEFINITION or CUSTOM_IMPORTS are present, explicitly instruct the coder to include them verbatim or use them as reference.
7. Do NOT write the Python code yourself. Write the PROMPT that asks for the code.
8. Ensure the tone is technical and precise.
Return ONLY the generated prompt text.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: systemPrompt,
});
return response.text.trim();
} catch (error) {
console.error("Prompt refinement failed:", error);
throw error; // Re-throw so caller can handle auth errors
}
};
export const validateArchitecture = async (nodes: Node<NodeData>[], edges: Edge[]): Promise<string> => {
const ai = getAiClient();
const graphRepresentation = {
nodes: nodes.map(n => ({
id: n.id,
type: n.data.type,
parameters: n.data.params
})),
edges: edges.map(e => ({
source: e.source,
target: e.target
}))
};
const prompt = `
Analyze this neural network architecture graph for validity.
Graph: ${JSON.stringify(graphRepresentation)}
Check for:
1. Shape mismatches (e.g., Conv2D output to Linear without Flatten).
2. Disconnected components.
3. Logical errors (e.g., MaxPool after Output).
4. Merge layer correctness (Concat/Add needs multiple inputs).
5. GenAI correctness (e.g., CrossAttention needs 2 inputs, VLM projection dims match).
Return a concise report. If valid, say "Architecture is valid.". If invalid, list specific errors and suggest fixes.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: prompt,
});
return response.text;
} catch (error) {
throw error;
}
}
/**
* Gets AI suggestions for improving the architecture.
*/
export const getArchitectureSuggestions = async (nodes: Node<NodeData>[], edges: Edge[]): Promise<string> => {
const ai = getAiClient();
const graphRepresentation = {
nodes: nodes.map(n => ({
id: n.id,
type: n.data.type,
parameters: n.data.params,
label: n.data.label
})),
edges: edges.map(e => ({
source: e.source,
target: e.target
}))
};
const prompt = `
You are a Senior Deep Learning Architect. Review the following neural network architecture graph.
Graph Structure:
${JSON.stringify(graphRepresentation, null, 2)}
Task: Provide 3 to 5 concrete, actionable suggestions to improve this model.
Focus on:
- Modern best practices (e.g., using LayerNorm vs BatchNorm in Transformers, SwiGLU vs ReLU).
- Architecture efficiency and parameter count optimization.
- Potential bottlenecks or vanishing gradient risks.
- Adding residuals or skip connections if the model is deep.
Format the output as a clean bulleted list. Keep it concise, professional and helpful.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: prompt,
});
return response.text;
} catch (error) {
throw error;
}
}
/**
* Implements the suggestions automatically.
*/
export const implementArchitectureSuggestions = async (
nodes: Node<NodeData>[],
edges: Edge[],
suggestions: string
): Promise<{ nodes: any[], edges: any[] }> => {
const ai = getAiClient();
const graphRepresentation = {
nodes: nodes.map(n => ({
id: n.id,
type: n.data.type,
parameters: n.data.params,
label: n.data.label,
position: n.position
})),
edges: edges.map(e => ({
source: e.source,
target: e.target
}))
};
const prompt = `
You are a Senior Implementation Engineer.
Task: Apply the following architectural suggestions to the provided graph JSON.
Current Graph:
${JSON.stringify(graphRepresentation)}
Suggestions to Implement:
"${suggestions}"
Instructions:
1. Modify the nodes and edges to incorporate the suggestions.
2. Maintain the layout (x, y positions) as best as possible, offsetting new nodes if added.
3. Ensure all LayerTypes are valid from the standard schema.
4. Return the complete, updated JSON with "nodes" and "edges" arrays.
5. Return ONLY raw JSON.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: prompt,
config: { responseMimeType: "application/json" }
});
const result = JSON.parse(response.text.trim());
return sanitizeGraph(result);
} catch (error) {
throw new Error("Failed to implement suggestions: " + (error instanceof Error ? error.message : String(error)));
}
};
/**
* Helper: Sanitizes the graph JSON from AI.
* Checks if 'type' is valid. If not, converts it to a CUSTOM layer to prevent crashes.
*/
const sanitizeGraph = (graphJson: { nodes: any[], edges: any[] }) => {
if (!graphJson || !graphJson.nodes) return graphJson;
graphJson.nodes = graphJson.nodes.map(node => {
// AI might return type in data.type or top-level type
let rawType = node.data?.type || node.type || 'Identity';
// Ensure rawType is a string
if (typeof rawType !== 'string') rawType = 'Identity';
// Check if this type exists in our known definitions
const isValid = Object.values(LayerType).includes(rawType as LayerType) && LAYER_DEFINITIONS[rawType as LayerType];
if (!isValid) {
console.warn(`Sanitizing unknown layer type: ${rawType}. Converting to CustomLayer.`);
return {
...node,
type: 'custom', // ReactFlow type
data: {
...node.data,
type: LayerType.CUSTOM,
label: node.data?.label || rawType,
params: {
...(node.data?.params || {}),
class_name: rawType, // Store original type name here
args: JSON.stringify(node.data?.params || {}).slice(0, 100) // Rough preserve of args
}
}
};
}
// Ensure standard structure for valid nodes
return {
...node,
type: 'custom',
data: {
...node.data,
type: rawType,
params: node.data?.params || {}
}
};
});
return graphJson;
};
/**
* Multi-agent graph generation.
* 1. Architect Agent: Drafts the layout.
* 2. Critic Agent: Reviews for errors.
* 3. Refiner Agent: Produces final JSON.
*/
export const generateGraphWithAgents = async (
userPrompt: string,
currentNodes: Node<NodeData>[] = [],
onStatusUpdate: (status: AgentStatus, log: string) => void
): Promise<{ nodes: any[], edges: any[] } | null> => {
const ai = getAiClient();
const layerSchema = Object.values(LAYER_DEFINITIONS).map(l => ({
type: l.type,
params: l.parameters.map(p => ({ name: p.name, type: p.type, options: p.options }))
}));
const schemaStr = JSON.stringify(layerSchema);
// --- Step 1: Architect ---
onStatusUpdate('architect', 'Architect is drafting initial layout...');
const context = currentNodes.length > 0
? `Current Graph Context: ${JSON.stringify(currentNodes.map(n => ({ id: n.id, type: n.data.type, label: n.data.label })))}`
: "Starting from scratch.";
const architectPrompt = `
Role: Senior Neural Network Architect.
Task: Create a preliminary graph layout (JSON) for the user request.
User Request: "${userPrompt}"
Context: ${context}
Available Layers: ${Object.keys(LAYER_DEFINITIONS).join(', ')}
Schema Reference: ${schemaStr}
Requirements:
1. Output valid JSON with "nodes" and "edges" arrays.
2. "nodes": { id, type='custom', position:{x,y}, data:{ type: LayerType, label: string, params: {} } }
3. Use correct LayerTypes from enum.
4. Layout nodes vertically (y+150 each step).
5. Connect edges logically.
6. If multi-input/output, arrange horizontally.
Return ONLY raw JSON.
`;
let draftJsonStr = "";
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: architectPrompt,
config: { responseMimeType: "application/json" }
});
draftJsonStr = response.text.trim();
} catch (e) {
throw e;
}
// --- Step 2: Critic ---
onStatusUpdate('critic', 'Critic is reviewing architecture for flaws...');
const criticPrompt = `
Role: Senior Lead Reviewer.
Task: Critique the following neural network architecture draft.
User Request: "${userPrompt}"
Draft Architecture: ${draftJsonStr}
Check for:
- Shape mismatches (e.g. 3D output into 2D input without flattening)
- Logical connection errors
- Missing essential layers (e.g. Activations, Normalization)
- Parameter errors (e.g. kernel size too large)
- Compliance with user request
Output a concise paragraph describing specific improvements needed. If perfect, say "No changes needed".
`;
let critique = "";
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: criticPrompt,
});
critique = response.text.trim();
} catch (e) {
console.warn("Critic agent failed, proceeding with draft.");
critique = "No critique available.";
}
// --- Step 3: Refiner ---
onStatusUpdate('refiner', 'Refiner is applying fixes and finalizing...');
const refinerPrompt = `
Role: Lead Engineer.
Task: Finalize the JSON architecture based on the critique.
Draft: ${draftJsonStr}
Critique: "${critique}"
Instructions:
1. Apply the fixes suggested in the critique.
2. Ensure the JSON structure is strictly { "nodes": [...], "edges": [...] }.
3. Ensure all node IDs are unique strings.
4. Ensure parameter values match the schema types.
5. Ensure "type" in top level node object is always 'custom'.
Return ONLY the final JSON.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: refinerPrompt,
config: { responseMimeType: "application/json" }
});
const finalJson = JSON.parse(response.text.trim());
onStatusUpdate('complete', 'Architecture built successfully!');
// SANITIZE: Prevent UI crashes by handling hallucinated types
return sanitizeGraph(finalJson);
} catch (e) {
throw new Error("Refiner agent failed to parse final JSON.");
}
};
/**
* Multi-agent code generation.
* Uses the detailed prompt to write, review, and polish PyTorch code.
*/
export const generateCodeWithAgents = async (
promptText: string,
onStatusUpdate: (status: AgentStatus, log: string) => void
): Promise<string> => {
const ai = getAiClient();
// --- Step 1: Coder (Architect) ---
onStatusUpdate('architect', 'Coder Agent is writing initial PyTorch implementation...');
const coderPrompt = `
Role: Senior Deep Learning Engineer.
Task: Write complete, runnable PyTorch code based on the following architecture prompt.
Prompt: "${promptText}"
Requirements:
- Use torch.nn.Module
- Include all necessary imports
- Handle forward pass logic exactly as described
- Include a 'if __name__ == "__main__":' block to test with dummy data
Return ONLY the Python code. No markdown formatting.
`;
let draftCode = "";
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: coderPrompt
});
draftCode = response.text.trim().replace(/```python/g, '').replace(/```/g, '');
} catch(e) {
throw e;
}
// --- Step 2: Reviewer (Critic) ---
onStatusUpdate('critic', 'Reviewer Agent is analyzing code for bugs and optimization...');
const reviewPrompt = `
Role: Code Reviewer.
Task: Review the following PyTorch code for errors, shape mismatches, or style issues.
Code:
${draftCode}
Original Prompt Request: "${promptText}"
Output a concise critique. If perfect, say "No changes needed".
`;
let critique = "";
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: reviewPrompt
});
critique = response.text.trim();
} catch(e) {
critique = "No critique available.";
}
// --- Step 3: Polisher (Refiner) ---
onStatusUpdate('refiner', 'Polisher Agent is finalizing the codebase...');
const polisherPrompt = `
Role: Senior Software Engineer.
Task: Refine the PyTorch code based on the critique.
Draft Code: ${draftCode}
Critique: ${critique}
Return ONLY the final Python code. No markdown formatting.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: polisherPrompt
});
let finalCode = response.text.trim().replace(/```python/g, '').replace(/```/g, '');
onStatusUpdate('complete', 'Code generation complete!');
return finalCode;
} catch(e) {
throw new Error("Polisher agent failed to generate code.");
}
};
/**
* Fix Architecture Errors with Agents (Debugger -> Architect -> Patcher)
*/
export const fixArchitectureErrors = async (
nodes: Node<NodeData>[],
edges: Edge[],
errorMsg: string,
onStatusUpdate: (status: AgentStatus, log: string) => void
): Promise<{ nodes: any[], edges: any[] } | null> => {
const ai = getAiClient();
const graphJson = JSON.stringify({
nodes: nodes.map(n => ({ id: n.id, type: n.data.type, label: n.data.label, params: n.data.params, position: n.position })),
edges: edges.map(e => ({ source: e.source, target: e.target }))
});
// --- Step 1: Debugger ---
onStatusUpdate('debugger', 'Debugger Agent is analyzing the error trace...');
const debuggerPrompt = `
Role: Senior Systems Debugger.
Task: Analyze the architecture graph and the reported error to pinpoint the root cause.
Graph: ${graphJson}
Error Message: "${errorMsg}"
Output a technical analysis of exactly what is wrong (e.g. "Node A connects to Node B but shapes [X] and [Y] are incompatible").
`;
let debugAnalysis = "";
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: debuggerPrompt,
});
debugAnalysis = response.text.trim();
} catch (e) {
throw e;
}
// --- Step 2: Architect ---
onStatusUpdate('architect', 'Architect Agent is planning the fix...');
const architectPrompt = `
Role: Solution Architect.
Task: Propose a specific fix for the identified issue.
Issue Analysis: ${debugAnalysis}
Instructions:
- Determine if nodes need to be added (e.g. Flatten, Reshape), removed, or reconnected.
- Determine if parameters need changing.
Output the plan in clear steps.
`;
let fixPlan = "";
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: architectPrompt,
});
fixPlan = response.text.trim();
} catch (e) {
fixPlan = "Apply necessary structural corrections.";
}
// --- Step 3: Patcher ---
onStatusUpdate('patcher', 'Patcher Agent is applying the fix to the graph...');
const patcherPrompt = `
Role: DevOps Engineer.
Task: Apply the fix to the graph JSON.
Current Graph: ${graphJson}
Fix Plan: ${fixPlan}
Requirements:
1. Return the complete, valid JSON with "nodes" and "edges".
2. Maintain existing node positions where possible, offset new nodes if added.
3. Ensure all LayerTypes are valid.
4. Return ONLY raw JSON.
`;
try {
const response = await ai.models.generateContent({
model: MODEL_NAME,
contents: patcherPrompt,
config: { responseMimeType: "application/json" }
});
const finalJson = JSON.parse(response.text.trim());
onStatusUpdate('complete', 'Fix applied successfully!');
return sanitizeGraph(finalJson);
} catch (e) {
throw new Error("Patcher agent failed to generate valid JSON.");
}
};