### What problem does this PR solve?
Hello, our use case requires LLM agent to invoke some tools, so I made a
simple implementation here.
This PR does two things:
1. A simple plugin mechanism based on `pluginlib`:
This mechanism lives in the `plugin` directory. It will only load
plugins from `plugin/embedded_plugins` for now.
A sample plugin `bad_calculator.py` is placed in
`plugin/embedded_plugins/llm_tools`, it accepts two numbers `a` and `b`,
then give a wrong result `a + b + 100`.
In the future, it can load plugins from external location with little
code change.
Plugins are divided into different types. The only plugin type supported
in this PR is `llm_tools`, which must implement the `LLMToolPlugin`
class in the `plugin/llm_tool_plugin.py`.
More plugin types can be added in the future.
2. A tool selector in the `Generate` component:
Added a tool selector to select one or more tools for LLM:
And with the `bad_calculator` tool, it results this with the `qwen-max`
model:
### Type of change
- [ ] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
Co-authored-by: Yingfeng <yingfeng.zhang@gmail.com>
### What problem does this PR solve?
Add code agent component.
#4977
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
This is a follow-up of #7088 , adding a knowledge base type input to the
`Begin` component, and a knowledge base selector to the agent flow debug
input panel:
then you can select one or more knowledge bases when testing the agent:
Note: the lines changed in `agent/component/retrieval.py` after line 94
are modified by `ruff format` from the `pre-commit` hooks, no functional
change.
### Type of change
- [ ] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
### What problem does this PR solve?
fixed errror when vars of cnt begin declare with key contain "begin"
_Briefly describe what this PR aims to solve. Include background context
that will help reviewers understand the purpose of the PR._
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Fix#6600
Hello, I have the same business requirement as #6600. My use case is:
We have many departments (> 20 now and increasing), and each department
has its own knowledge base. Because the agent workflow is the same, so I
want to change the knowledge base on the fly, instead of creating agents
for every department.
It now looks like this:
Knowledge bases can be selected from the dropdown, and passed through
the variables in the table. All selected knowledge bases are used for
retrieval.
### Type of change
- [ ] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
### What problem does this PR solve?
Some models force thinking, resulting in the absence of the think tag in
the returned content
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Fix the issue where waiting tasks couldn't be processed when upstream
components were "switch", "categorize", or "relevant" and the normal
processing path couldn't continue.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [ ] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
### What problem does this PR solve?
Actually fix#6241
Hello, I ran into the same problem as #6241. When I'm testing my agent
flow in the web ui using `Run` button with a file input, the retrieval
component always gave an empty output.
In the code I found that:
`web/src/pages/flow/debug-content/index.tsx`:
```tsx
const onOk = useCallback(async () => {
const values = await form.validateFields();
const nextValues = Object.entries(values).map(([key, value]) => {
const item = parameters[Number(key)];
let nextValue = value;
if (Array.isArray(value)) {
nextValue = ``;
value.forEach((x) => {
nextValue +=
x?.originFileObj instanceof File
? `${x.name}\n${x.response?.data}\n----\n` // Here, the file content always ends in '\n'
: `${x.url}\n${x.result}\n----\n`;
});
}
return { ...item, value: nextValue };
});
ok(nextValues);
}, [form, ok, parameters]);
```
while in the `agent/component/retrieval.py`:
```python
def _run(self, history, **kwargs):
query = self.get_input()
query = str(query["content"][0]) if "content" in query else ""
lines = query.split('\n') # inputs are split to ['xxx','yyy','----','']
query = lines[-1] if lines else "" # Here we always get '', thus no result
kbs = KnowledgebaseService.get_by_ids(self._param.kb_ids)
if not kbs:
return Retrieval.be_output("")
```
so the code will never got correct result.
I'm not sure why the input needs such a split here, so I just removed
the splitting, and it works well on my side.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [ ] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
### What problem does this PR solve?
When I use the categorization operator, I find that if the keyword I
want to Categorize appears repeatedly in the input, then I cannot judge
the word that appears most frequently. Instead, I simply get the word
that matches and return all the ones that have made the following
changes to the categorize filter.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
- [x] Refactoring
- [x] Performance Improvement
### What problem does this PR solve?
For now, if you use thinking model (deepseek-r1:32b with ollama server
in my case) in "Keyword" node, result contains all <think> block and so
node return not only keywords
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [ ] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):
###Address Problem:
The original implementation used re.sub(r"(\\\"|\")", "", content) which
stripped all quotes from the processed content. While this worked for
simple Jinja2-rendered templates, it caused formatting issues when :
-Quotes were required in the final output (e.g., JSON, Python Code
strings)
###Solution:
1. Selective JSON Serialization.
2. Removed Global Quote Removal
### What problem does this PR solve?
This PR addresses an issue in template processing where all quotation
marks (" and \") were being removed from content, potentially corrupting
string formatting in rendered outputs. **In fact, extra quotes is
generated by json.dumps(v, ensure_ascii=False).**
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Fix: Resolved a bug where sibling components in Canvas were not
restricted to fetching data from the upstream when parallel components
were present.
Issue: When parallel components existed in Canvas, sibling components
incorrectly fetched data without being limited to the upstream scope,
causing data retrieval issues.
Solution: Adjusted the data fetching logic to ensure sibling components
only retrieve data from the upstream scope.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
**generate.py 更新：**
问题：部分模型提供商对输入对话内容的格式有严格校验，要求第一条内容的 role 不能为 assistant，否则会报错。
解决：删除了系统设置的 agent 开场白，确保传递给模型的对话内容中，第一条内容的 role 不为 assistant。
**retrieval.py 更新：**
问题：当前知识库检索使用全部对话内容作为输入，可能导致检索结果不准确。
解决：改为仅使用用户最后提出的一个问题进行知识库检索，提高检索的准确性。
**Update generate.py:**
Issue: Some model providers have strict validation rules for the format
of input conversation content, requiring that the role of the first
content must not be assistant. Otherwise, an error will occur.
Solution: Removed the system-set agent opening statement to ensure that
the role of the first content in the conversation passed to the model is
not assistant.
**Update retrieval.py:**
Issue: The current knowledge base retrieval uses the entire conversation
content as input, which may lead to inaccurate retrieval results.
Solution: Changed the retrieval logic to use only the last question
asked by the user for knowledge base retrieval, improving retrieval
accuracy.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] Performance Improvement
### What problem does this PR solve?
```
Invoke agent
To be able to interact dynamically with the API, there is a customizable Data Type JSON or FormData, the default is JSON
```
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
---------
Co-authored-by: Kevin Hu <kevinhu.sh@gmail.com>
### What problem does this PR solve?
seems no need use ABC here, there's no `abstractmethod` here
### Type of change
- [x] Performance Improvement
Signed-off-by: yihong0618 <zouzou0208@gmail.com>
### What problem does this PR solve?
Invoke component can be used to call third party services.
Tried GET/POST/PUT from web UI, and found PUT request failed like this:
(test api: api/v1/chats/<assistant_id>)
```
{"code":100,"data":null,"message":"AttributeError("'NoneType' object has
no attribute 'get'")"}
```
Root cause: Invoke PUT with a 'data=args' parameter, which is a form-encoded data, however the default content type setting of request header is application/json. The test api could not deal with such case.
Fix: use the 'json' parameter of reqeusts.put(), same as Invoke POST. Do not use the 'data' parameter.
Another way is to use 'data=json.dumps(args)'.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
