Added new tools
Browse files- tools/tools_on_modal_labs.py +280 -0
tools/tools_on_modal_labs.py
ADDED
@@ -0,0 +1,280 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
This module provides tools for searching and retrieving context from a knowledge base,
|
3 |
+
and for conducting a research workflow that includes searching, writing, and reviewing reports.
|
4 |
+
The tools are designed to be used with Modal Labs for scalable and efficient processing.
|
5 |
+
The technology stack includes FastAPI for the API interface, GroundX for knowledge base search,
|
6 |
+
LlamaIndex for LLM workflows, Nebius for LLM, and Modal Labs for tool execution.
|
7 |
+
"""
|
8 |
+
|
9 |
+
import os
|
10 |
+
import asyncio
|
11 |
+
|
12 |
+
import modal
|
13 |
+
from pydantic import BaseModel
|
14 |
+
|
15 |
+
image = modal.Image.debian_slim().pip_install(
|
16 |
+
"fastapi[standard]",
|
17 |
+
"groundx",
|
18 |
+
"llama-index",
|
19 |
+
"llama-index-llms-nebius",
|
20 |
+
"duckduckgo-search",
|
21 |
+
"langchain-community")
|
22 |
+
|
23 |
+
app = modal.App(name="hackathon-mcp-tools", image=image)
|
24 |
+
|
25 |
+
class QueryInput(BaseModel):
|
26 |
+
query: str
|
27 |
+
|
28 |
+
@app.function(secrets=[
|
29 |
+
modal.Secret.from_name("hackathon-secret", required_keys=["GROUNDX_API_KEY"])
|
30 |
+
])
|
31 |
+
@modal.fastapi_endpoint(docs=True, method="POST")
|
32 |
+
def search_rag_context(queryInput: QueryInput) -> str:
|
33 |
+
"""
|
34 |
+
Searches and retrieves relevant context from a knowledge base,
|
35 |
+
based on the user's query.
|
36 |
+
Args:
|
37 |
+
query: The search query supplied by the user.
|
38 |
+
Returns:
|
39 |
+
str: Relevant text content that can be used by the LLM to answer the query.
|
40 |
+
"""
|
41 |
+
|
42 |
+
result = search_groundx_for_rag_context(queryInput.query)
|
43 |
+
|
44 |
+
print("\n\n=============================")
|
45 |
+
print(f"RAG Search Result: {result}")
|
46 |
+
print("=============================\n")
|
47 |
+
|
48 |
+
return
|
49 |
+
|
50 |
+
def search_groundx_for_rag_context(query) -> str:
|
51 |
+
from groundx import GroundX
|
52 |
+
|
53 |
+
client = GroundX(api_key=os.getenv("GROUNDX_API_KEY") or '')
|
54 |
+
response = client.search.content(
|
55 |
+
id=os.getenv("GROUNDX_BUCKET_ID"),
|
56 |
+
query=query,
|
57 |
+
n=10,
|
58 |
+
)
|
59 |
+
|
60 |
+
return response.search.text or "No relevant context found"
|
61 |
+
|
62 |
+
from llama_index.llms.nebius import NebiusLLM
|
63 |
+
|
64 |
+
# llama-index workflow classes
|
65 |
+
from llama_index.core.workflow import Context
|
66 |
+
from llama_index.core.agent.workflow import (
|
67 |
+
FunctionAgent,
|
68 |
+
AgentWorkflow,
|
69 |
+
AgentOutput,
|
70 |
+
ToolCall,
|
71 |
+
ToolCallResult,
|
72 |
+
)
|
73 |
+
|
74 |
+
from langchain.utilities import DuckDuckGoSearchAPIWrapper
|
75 |
+
|
76 |
+
@app.function(secrets=[
|
77 |
+
modal.Secret.from_name("hackathon-secret", required_keys=["NEBIUS_API_KEY", "AGENT_MODEL"])
|
78 |
+
])
|
79 |
+
@modal.fastapi_endpoint(docs=True, method="POST")
|
80 |
+
def run_research_workflow(queryInput: QueryInput) -> str:
|
81 |
+
handler = asyncio.run(execute_research_workflow(queryInput.query))
|
82 |
+
result = asyncio.run(final_report(handler))
|
83 |
+
return result
|
84 |
+
|
85 |
+
NEBIUS_API_KEY = os.getenv("NEBIUS_API_KEY")
|
86 |
+
AGENT_MODEL = os.getenv("AGENT_MODEL", "meta-llama/Meta-Llama-3.1-8B-Instruct")
|
87 |
+
|
88 |
+
# Load an LLM
|
89 |
+
llm = NebiusLLM(
|
90 |
+
api_key=NEBIUS_API_KEY,
|
91 |
+
model=AGENT_MODEL,
|
92 |
+
is_function_calling_model=True
|
93 |
+
)
|
94 |
+
|
95 |
+
# Search tools using DuckDuckGo
|
96 |
+
duckduckgo = DuckDuckGoSearchAPIWrapper()
|
97 |
+
|
98 |
+
MAX_SEARCH_CALLS = 2 # Limit the number of searches to 2
|
99 |
+
search_call_count = 0
|
100 |
+
past_queries = set()
|
101 |
+
|
102 |
+
async def duckduckgo_search(query: str) -> str:
|
103 |
+
"""
|
104 |
+
A DuckDuckGo-based search limiting number of searches and avoiding duplicates.
|
105 |
+
"""
|
106 |
+
global search_call_count, past_queries
|
107 |
+
|
108 |
+
# Check for duplicate queries
|
109 |
+
if query in past_queries:
|
110 |
+
return f"Already searched for '{query}'."
|
111 |
+
|
112 |
+
# Check if we've reached the max search calls
|
113 |
+
if search_call_count >= MAX_SEARCH_CALLS:
|
114 |
+
return "Search limit reached."
|
115 |
+
|
116 |
+
# Otherwise, perform the search
|
117 |
+
search_call_count += 1
|
118 |
+
past_queries.add(query)
|
119 |
+
|
120 |
+
result = duckduckgo.run(query)
|
121 |
+
return str(result)
|
122 |
+
|
123 |
+
# Research tools
|
124 |
+
async def save_research(ctx: Context, notes: str, notes_title: str) -> str:
|
125 |
+
"""
|
126 |
+
Store research notes under a given title in the shared context.
|
127 |
+
"""
|
128 |
+
|
129 |
+
current_state = await ctx.get("state")
|
130 |
+
if "research_notes" not in current_state:
|
131 |
+
current_state["research_notes"] = {}
|
132 |
+
current_state["research_notes"][notes_title] = notes
|
133 |
+
await ctx.set("state", current_state)
|
134 |
+
return "Notes saved."
|
135 |
+
|
136 |
+
# Report tools
|
137 |
+
async def write_report(ctx: Context, report_content: str) -> str:
|
138 |
+
"""
|
139 |
+
Write a report in markdown, storing it in the shared context.
|
140 |
+
"""
|
141 |
+
|
142 |
+
current_state = await ctx.get("state")
|
143 |
+
current_state["report_content"] = report_content
|
144 |
+
await ctx.set("state", current_state)
|
145 |
+
return "Report written."
|
146 |
+
|
147 |
+
# Review tools
|
148 |
+
async def review_report(ctx: Context, review: str) -> str:
|
149 |
+
"""
|
150 |
+
Review the report and store feedback in the shared context.
|
151 |
+
"""
|
152 |
+
|
153 |
+
current_state = await ctx.get("state")
|
154 |
+
current_state["review"] = review
|
155 |
+
await ctx.set("state", current_state)
|
156 |
+
return "Report reviewed."
|
157 |
+
|
158 |
+
|
159 |
+
# We have three agents with distinct responsibilities:
|
160 |
+
# - The ResearchAgent is responsible for gathering information from the web.
|
161 |
+
# - The WriteAgent is responsible for writing the report.
|
162 |
+
# - The ReviewAgent is responsible for reviewing the report.
|
163 |
+
|
164 |
+
# The ResearchAgent uses the DuckDuckGoSearchAPIWrapper to search the web.
|
165 |
+
|
166 |
+
research_agent = FunctionAgent(
|
167 |
+
name="ResearchAgent",
|
168 |
+
description=(
|
169 |
+
"A research agent that searches the web using Google search through SerpAPI. "
|
170 |
+
"It must not exceed 2 searches total, and must avoid repeating the same query. "
|
171 |
+
"Once sufficient information is collected, it should hand off to the WriteAgent."
|
172 |
+
),
|
173 |
+
system_prompt=(
|
174 |
+
"You are the ResearchAgent. Your goal is to gather sufficient information on the topic. "
|
175 |
+
"Only perform at most 2 distinct searches. If you have enough information or have reached 2 searches, "
|
176 |
+
"handoff to the WriteAgent. Avoid infinite loops! If search throws an error, stop further work and skip WriteAgent and ReviewAgent and return."
|
177 |
+
"Respect invocation limits and cooldown periods."
|
178 |
+
),
|
179 |
+
llm=llm,
|
180 |
+
tools=[
|
181 |
+
duckduckgo_search,
|
182 |
+
save_research
|
183 |
+
],
|
184 |
+
max_iterations=2, # Limit to 2 iterations to prevent infinite loops
|
185 |
+
cooldown=5, # Cooldown to prevent rapid re-querying
|
186 |
+
can_handoff_to=["WriteAgent"]
|
187 |
+
)
|
188 |
+
|
189 |
+
write_agent = FunctionAgent(
|
190 |
+
name="WriteAgent",
|
191 |
+
description=(
|
192 |
+
"Writes a markdown report based on the research notes. "
|
193 |
+
"Then hands off to the ReviewAgent for feedback."
|
194 |
+
),
|
195 |
+
system_prompt=(
|
196 |
+
"You are the WriteAgent. Draft a structured markdown report based on the notes. "
|
197 |
+
"If there is no report content or research notes, stop further work and skip ReviewAgent."
|
198 |
+
"Do not attempt more than one write attempt. "
|
199 |
+
"After writing, hand off to the ReviewAgent."
|
200 |
+
"Respect invocation limits and cooldown periods."
|
201 |
+
),
|
202 |
+
llm=llm,
|
203 |
+
tools=[write_report],
|
204 |
+
max_iterations=2, # Limit to 2 iterations to prevent infinite loops
|
205 |
+
cooldown=5, # Cooldown to prevent rapid re-querying
|
206 |
+
can_handoff_to=["ReviewAgent", "ResearchAgent"]
|
207 |
+
)
|
208 |
+
|
209 |
+
review_agent = FunctionAgent(
|
210 |
+
name="ReviewAgent",
|
211 |
+
description=(
|
212 |
+
"Reviews the final report for correctness. Approves or requests changes."
|
213 |
+
),
|
214 |
+
system_prompt=(
|
215 |
+
"You are the ReviewAgent. If there is no research notes or report content, skip this step and return."
|
216 |
+
"Do not attempt more than one review attempt. "
|
217 |
+
"Read the report, provide feedback, and either approve "
|
218 |
+
"or request revisions. If revisions are needed, handoff to WriteAgent."
|
219 |
+
"Respect invocation limits and cooldown periods."
|
220 |
+
),
|
221 |
+
llm=llm,
|
222 |
+
tools=[review_report],
|
223 |
+
max_iterations=2, # Limit to 2 iterations to prevent infinite loops
|
224 |
+
cooldown=5, # Cooldown to prevent rapid re-querying
|
225 |
+
can_handoff_to=["WriteAgent"]
|
226 |
+
)
|
227 |
+
|
228 |
+
agent_workflow = AgentWorkflow(
|
229 |
+
agents=[research_agent, write_agent, review_agent],
|
230 |
+
root_agent=research_agent.name, # Start with the ResearchAgent
|
231 |
+
initial_state={
|
232 |
+
"research_notes": {},
|
233 |
+
"report_content": "Not written yet.",
|
234 |
+
"review": "Review required.",
|
235 |
+
},
|
236 |
+
)
|
237 |
+
|
238 |
+
async def execute_research_workflow(query: str):
|
239 |
+
handler = agent_workflow.run(
|
240 |
+
user_msg=(
|
241 |
+
query
|
242 |
+
)
|
243 |
+
)
|
244 |
+
|
245 |
+
current_agent = None
|
246 |
+
|
247 |
+
async for event in handler.stream_events():
|
248 |
+
if hasattr(event, "current_agent_name") and event.current_agent_name != current_agent:
|
249 |
+
current_agent = event.current_agent_name
|
250 |
+
print(f"\n{'='*50}")
|
251 |
+
print(f"π€ Agent: {current_agent}")
|
252 |
+
print(f"{'='*50}\n")
|
253 |
+
|
254 |
+
# Print outputs or tool calls
|
255 |
+
if isinstance(event, AgentOutput):
|
256 |
+
if event.response.content:
|
257 |
+
print("π€ Output:", event.response.content)
|
258 |
+
if event.tool_calls:
|
259 |
+
print("π οΈ Planning to use tools:", [call.tool_name for call in event.tool_calls])
|
260 |
+
|
261 |
+
elif isinstance(event, ToolCall):
|
262 |
+
print(f"π¨ Calling Tool: {event.tool_name}")
|
263 |
+
print(f" With arguments: {event.tool_kwargs}")
|
264 |
+
|
265 |
+
elif isinstance(event, ToolCallResult):
|
266 |
+
print(f"π§ Tool Result ({event.tool_name}):")
|
267 |
+
print(f" Arguments: {event.tool_kwargs}")
|
268 |
+
print(f" Output: {event.tool_output}")
|
269 |
+
|
270 |
+
return handler
|
271 |
+
|
272 |
+
async def final_report(handler) -> str:
|
273 |
+
"""Retrieve the final report from the context."""
|
274 |
+
final_state = await handler.ctx.get("state")
|
275 |
+
print("\n\n=============================")
|
276 |
+
print("FINAL REPORT:\n")
|
277 |
+
print(final_state["report_content"])
|
278 |
+
print("=============================\n")
|
279 |
+
|
280 |
+
return final_state["report_content"]
|