changes
This commit is contained in:
+31
-13
@@ -4,7 +4,7 @@ import ollama from '@/lib/ollama';
|
||||
import { allTools, executeTool, getRegisteredTools } from '@/lib/tools';
|
||||
|
||||
// Maximum number of tool call iterations to prevent infinite loops
|
||||
const MAX_TOOL_ITERATIONS = 10;
|
||||
const MAX_TOOL_ITERATIONS = 5;
|
||||
|
||||
// Maximum tokens for model responses (effectively unlimited for long reasoning)
|
||||
const MAX_RESPONSE_TOKENS = 32768;
|
||||
@@ -12,22 +12,30 @@ const MAX_RESPONSE_TOKENS = 32768;
|
||||
// System prompt for tool-enabled conversations
|
||||
const TOOL_SYSTEM_PROMPT = `You are a helpful AI assistant with access to various tools.
|
||||
|
||||
## Tool Usage Guidelines
|
||||
## CRITICAL Tool Usage Rules
|
||||
|
||||
1. **Think before acting**: Before using a tool, briefly consider what information you need and which tool(s) would best provide it.
|
||||
1. **ONE attempt per tool**: If a tool returns a result (even partial), USE that result. Do NOT retry the same tool with different parameters.
|
||||
|
||||
2. **Be specific**: When a query is vague (e.g., "weather in Arizona"), ask for clarification OR make a reasonable specific choice (e.g., Phoenix, the capital). Don't call multiple tools for different interpretations.
|
||||
2. **Accept first valid result**: When you get weather, search results, or any data - use it immediately in your response. Don't try to "improve" it.
|
||||
|
||||
3. **Minimize tool calls**: Use the minimum number of tool calls needed. For example:
|
||||
- If asked about weather in "Paris", call once for Paris, France (the most likely intent)
|
||||
- If asked about "3 cities", call exactly 3 times
|
||||
- Don't call the same tool repeatedly with slight variations
|
||||
3. **Handle errors gracefully**: If a tool fails, tell the user. Do NOT retry with variations.
|
||||
|
||||
4. **Handle errors gracefully**: If a tool call fails, explain the issue to the user rather than retrying with random variations.
|
||||
4. **Give a final answer**: After using tools, you MUST provide a helpful summary to the user. Never end with just tool calls.
|
||||
|
||||
5. **Synthesize results**: After receiving tool results, provide a clear, helpful summary rather than just echoing the raw data.
|
||||
5. **Be specific with locations**: For vague locations like "Arizona", pick the most likely city (e.g., Phoenix) and use it once.
|
||||
|
||||
When using thinking/reasoning models: Use your <think> block to plan which tools to use and why, then execute efficiently.`;
|
||||
Example of CORRECT behavior:
|
||||
- User: "Weather in Arizona"
|
||||
- You: Call get_weather with "Phoenix, Arizona" ONCE
|
||||
- Tool returns result
|
||||
- You: Summarize the weather for the user
|
||||
|
||||
Example of WRONG behavior:
|
||||
- Calling the same tool multiple times with slight variations
|
||||
- Not providing a final response after getting tool results`;
|
||||
|
||||
// Message to inject when max iterations reached to force a response
|
||||
const FORCE_RESPONSE_MESSAGE = `You have made several tool calls. Now you MUST provide a final response to the user based on the information you have gathered. Do NOT make any more tool calls. Summarize what you learned and answer the user's question.`;
|
||||
|
||||
/**
|
||||
* Parse text-based tool calls from model output
|
||||
@@ -120,12 +128,22 @@ export async function POST(request: NextRequest) {
|
||||
while (iterations < MAX_TOOL_ITERATIONS) {
|
||||
iterations++;
|
||||
|
||||
// Call Ollama with tools if enabled
|
||||
// If we're on the last allowed iteration, force the model to respond without tools
|
||||
const isLastIteration = iterations === MAX_TOOL_ITERATIONS;
|
||||
if (isLastIteration) {
|
||||
// Add a message forcing the model to give a final response
|
||||
workingMessages.push({
|
||||
role: 'user',
|
||||
content: FORCE_RESPONSE_MESSAGE,
|
||||
});
|
||||
}
|
||||
|
||||
// Call Ollama with tools if enabled (but disable on last iteration)
|
||||
// Enable thinking mode for models that support it (like deepseek-r1)
|
||||
const response = await ollama.chat({
|
||||
model,
|
||||
messages: workingMessages,
|
||||
tools: enableTools ? allTools : undefined,
|
||||
tools: enableTools && !isLastIteration ? allTools : undefined,
|
||||
stream: true,
|
||||
think: true, // Enable thinking mode - model will separate thinking from response
|
||||
options: {
|
||||
|
||||
Reference in New Issue
Block a user