agentos-railway-template/agents/web_search.py at main · agno-agi/agentos-railway-template · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
"""
WebSearch Agent
===============
"""

from os import getenv

from agno.agent import Agent
from agno.tools.mcp import MCPTools
from agno.tools.parallel import ParallelTools

from app.settings import default_model
from db import get_postgres_db

# When PARALLEL_API_KEY is set, use the official parallel-web SDK —
# the agent gets `parallel_search` and `parallel_extract` directly.
# Without a key, fall back to the keyless MCP endpoint and the agent
# gets `web_search` and `web_fetch` instead. AgentOS handles MCP
# connect/close as part of its lifespan.
if getenv("PARALLEL_API_KEY"):
    web_tools: ParallelTools | MCPTools = ParallelTools()
else:
    web_tools = MCPTools(url="https://search.parallel.ai/mcp", transport="streamable-http")


WEB_SEARCH_INSTRUCTIONS = """\
Search the web for current information.

Workflow:
1. Use the search tool to find candidate sources for the question.
2. For recent-event, “latest,” or “recently” questions, answer only from search results you actually found in this run; do not infer newer publications, titles, or dates beyond what the results support.
3. When the user asks about specific pages, or when search snippets are too thin to safely summarize a recent claim, follow up with the extract / fetch tool to read the most relevant URLs before answering.
4. Cite the sources you used as plain URLs. Prefer recent, authoritative pages. If you cannot find a good answer, say so plainly.
"""


web_search = Agent(
    id="web-search",
    name="WebSearch",
    model=default_model(),
    db=get_postgres_db(),
    tools=[web_tools],
    instructions=WEB_SEARCH_INSTRUCTIONS,
    enable_agentic_memory=True,
    add_datetime_to_context=True,
    add_history_to_context=True,
    num_history_runs=5,
    markdown=True,
)