Martin Bär
commited on
Commit
·
09a402e
1
Parent(s):
ba66f78
Only activate langfuse conditionally
Browse files- .gitignore +1 -1
- basic_agent.py +18 -8
.gitignore
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
__pycache__
|
| 2 |
.venv
|
| 3 |
-
|
| 4 |
.env
|
|
|
|
| 1 |
__pycache__
|
| 2 |
.venv
|
| 3 |
+
*.ipynb
|
| 4 |
.env
|
basic_agent.py
CHANGED
|
@@ -8,14 +8,20 @@ from llama_index.core.tools.tool_spec.load_and_search import LoadAndSearchToolSp
|
|
| 8 |
from llama_index.readers.web import SimpleWebPageReader
|
| 9 |
from llama_index.core.tools.ondemand_loader_tool import OnDemandLoaderTool
|
| 10 |
from langfuse.llama_index import LlamaIndexInstrumentor
|
| 11 |
-
|
| 12 |
-
# Langfuse
|
| 13 |
-
instrumentor = LlamaIndexInstrumentor()
|
| 14 |
-
instrumentor.start()
|
| 15 |
|
| 16 |
class BasicAgent:
|
| 17 |
-
def __init__(self):
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
# Initialize tools
|
| 21 |
tool_spec = DuckDuckGoSearchToolSpec()
|
|
@@ -31,6 +37,8 @@ class BasicAgent:
|
|
| 31 |
|
| 32 |
# Convert into a LoadAndSearchToolSpec because the wikipedia search tool returns
|
| 33 |
# entire Wikipedia pages and this can pollute the context window of the LLM
|
|
|
|
|
|
|
| 34 |
wiki_search_tool_las = LoadAndSearchToolSpec.from_defaults(wiki_search_tool).to_tool_list()
|
| 35 |
|
| 36 |
webpage_tool = OnDemandLoaderTool.from_defaults(
|
|
@@ -41,7 +49,7 @@ class BasicAgent:
|
|
| 41 |
|
| 42 |
# Create Alfred with all the tools
|
| 43 |
self.agent = AgentWorkflow.from_tools_or_functions(
|
| 44 |
-
wiki_search_tool_las + [search_tool], # webpage_tool does not work properly
|
| 45 |
llm=llm,
|
| 46 |
verbose=True,
|
| 47 |
system_prompt=("You are a helpful agent that can search Wikipedia and the Internet for answers. "
|
|
@@ -52,5 +60,7 @@ class BasicAgent:
|
|
| 52 |
|
| 53 |
async def __call__(self, question: str) -> str:
|
| 54 |
response = await self.agent.run(user_msg=question) # ctx=self.ctx)
|
| 55 |
-
|
|
|
|
|
|
|
| 56 |
return response.response.content
|
|
|
|
| 8 |
from llama_index.readers.web import SimpleWebPageReader
|
| 9 |
from llama_index.core.tools.ondemand_loader_tool import OnDemandLoaderTool
|
| 10 |
from langfuse.llama_index import LlamaIndexInstrumentor
|
| 11 |
+
from llama_index.llms.ollama import Ollama
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
class BasicAgent:
|
| 14 |
+
def __init__(self, ollama=False, langfuse=True):
|
| 15 |
+
if not ollama:
|
| 16 |
+
llm = HuggingFaceInferenceAPI(model_name="Qwen/Qwen2.5-Coder-32B-Instruct")
|
| 17 |
+
else:
|
| 18 |
+
llm = Ollama(model="mistral:latest", request_timeout=120.0)
|
| 19 |
+
|
| 20 |
+
# Langfuse
|
| 21 |
+
self.langfuse = langfuse
|
| 22 |
+
if self.langfuse:
|
| 23 |
+
self.instrumentor = LlamaIndexInstrumentor()
|
| 24 |
+
self.instrumentor.start()
|
| 25 |
|
| 26 |
# Initialize tools
|
| 27 |
tool_spec = DuckDuckGoSearchToolSpec()
|
|
|
|
| 37 |
|
| 38 |
# Convert into a LoadAndSearchToolSpec because the wikipedia search tool returns
|
| 39 |
# entire Wikipedia pages and this can pollute the context window of the LLM
|
| 40 |
+
|
| 41 |
+
# TODO this does not work so well. We need to make the retriever return the top 5 chunks or sth.
|
| 42 |
wiki_search_tool_las = LoadAndSearchToolSpec.from_defaults(wiki_search_tool).to_tool_list()
|
| 43 |
|
| 44 |
webpage_tool = OnDemandLoaderTool.from_defaults(
|
|
|
|
| 49 |
|
| 50 |
# Create Alfred with all the tools
|
| 51 |
self.agent = AgentWorkflow.from_tools_or_functions(
|
| 52 |
+
wiki_search_tool_las + [search_tool], # webpage_tool does not work properly - cookies etc
|
| 53 |
llm=llm,
|
| 54 |
verbose=True,
|
| 55 |
system_prompt=("You are a helpful agent that can search Wikipedia and the Internet for answers. "
|
|
|
|
| 60 |
|
| 61 |
async def __call__(self, question: str) -> str:
|
| 62 |
response = await self.agent.run(user_msg=question) # ctx=self.ctx)
|
| 63 |
+
|
| 64 |
+
if self.langfuse:
|
| 65 |
+
self.instrumentor.flush()
|
| 66 |
return response.response.content
|