### What problem does this PR solve?
[https://github.com/infiniflow/ragflow/issues/8324](url)
docker image version: v0.19.1
The `_clean_conf` function was not implemented in the `_chat` and
`chat_streamly` methods of the `GeminiChat` class, causing the error
"Unknown field for GenerationConfig: max_tokens" when the default LLM
config includes the "max_tokens" parameter.
**Buggy Code(ragflow/rag/llm/chat_model.py)**
```python
class GeminiChat(Base):
def __init__(self, key, model_name, base_url=None, **kwargs):
super().__init__(key, model_name, base_url=base_url, **kwargs)
from google.generativeai import GenerativeModel, client
client.configure(api_key=key)
_client = client.get_default_generative_client()
self.model_name = "models/" + model_name
self.model = GenerativeModel(model_name=self.model_name)
self.model._client = _client
def _clean_conf(self, gen_conf):
for k in list(gen_conf.keys()):
if k not in ["temperature", "top_p"]:
del gen_conf[k]
return gen_conf
def _chat(self, history, gen_conf):
from google.generativeai.types import content_types
system = history[0]["content"] if history and history[0]["role"] == "system" else ""
hist = []
for item in history:
if item["role"] == "system":
continue
hist.append(deepcopy(item))
item = hist[-1]
if "role" in item and item["role"] == "assistant":
item["role"] = "model"
if "role" in item and item["role"] == "system":
item["role"] = "user"
if "content" in item:
item["parts"] = item.pop("content")
if system:
self.model._system_instruction = content_types.to_content(system)
response = self.model.generate_content(hist, generation_config=gen_conf)
ans = response.text
return ans, response.usage_metadata.total_token_count
def chat_streamly(self, system, history, gen_conf):
from google.generativeai.types import content_types
if system:
self.model._system_instruction = content_types.to_content(system)
#❌_clean_conf was not implemented
for k in list(gen_conf.keys()):
if k not in ["temperature", "top_p", "max_tokens"]:
del gen_conf[k]
for item in history:
if "role" in item and item["role"] == "assistant":
item["role"] = "model"
if "content" in item:
item["parts"] = item.pop("content")
ans = ""
try:
response = self.model.generate_content(history, generation_config=gen_conf, stream=True)
for resp in response:
ans = resp.text
yield ans
yield response._chunks[-1].usage_metadata.total_token_count
except Exception as e:
yield ans + "\n**ERROR**: " + str(e)
yield 0
```
**Implement the _clean_conf function**
```python
class GeminiChat(Base):
def __init__(self, key, model_name, base_url=None, **kwargs):
super().__init__(key, model_name, base_url=base_url, **kwargs)
from google.generativeai import GenerativeModel, client
client.configure(api_key=key)
_client = client.get_default_generative_client()
self.model_name = "models/" + model_name
self.model = GenerativeModel(model_name=self.model_name)
self.model._client = _client
def _clean_conf(self, gen_conf):
for k in list(gen_conf.keys()):
if k not in ["temperature", "top_p"]:
del gen_conf[k]
return gen_conf
def _chat(self, history, gen_conf):
from google.generativeai.types import content_types
#✅ implement _clean_conf to remove the wrong parameters
gen_conf = self._clean_conf(gen_conf)
system = history[0]["content"] if history and history[0]["role"] == "system" else ""
hist = []
for item in history:
if item["role"] == "system":
continue
hist.append(deepcopy(item))
item = hist[-1]
if "role" in item and item["role"] == "assistant":
item["role"] = "model"
if "role" in item and item["role"] == "system":
item["role"] = "user"
if "content" in item:
item["parts"] = item.pop("content")
if system:
self.model._system_instruction = content_types.to_content(system)
response = self.model.generate_content(hist, generation_config=gen_conf)
ans = response.text
return ans, response.usage_metadata.total_token_count
def chat_streamly(self, system, history, gen_conf):
from google.generativeai.types import content_types
#✅ implement _clean_conf to remove the wrong parameters
gen_conf = self._clean_conf(gen_conf)
if system:
self.model._system_instruction = content_types.to_content(system)
#✅Removed duplicate parameter filtering logic "for k in list(gen_conf.keys()):"
for item in history:
if "role" in item and item["role"] == "assistant":
item["role"] = "model"
if "content" in item:
item["parts"] = item.pop("content")
ans = ""
try:
response = self.model.generate_content(history, generation_config=gen_conf, stream=True)
for resp in response:
ans = resp.text
yield ans
yield response._chunks[-1].usage_metadata.total_token_count
except Exception as e:
yield ans + "\n**ERROR**: " + str(e)
yield 0
```
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
---------
Co-authored-by: Kevin Hu <kevinhu.sh@gmail.com>
### What problem does this PR solve?
image_version: v0.19.1
This PR fixes a bug in the HuggingFaceEmBedding API method that was
causing AssertionError: assert len(vects) == len(docs) during the
document embedding process.
#### Problem
The HuggingFaceEmbed.encode() method had an early return statement
inside the for loop, causing it to return after processing only the
first text input instead of processing all texts in the input list.
**Error Messenge**
```python
AssertionError: assert len(vects) == len(docs) # input chunks != embedded vectors from embedding api
File "/ragflow/rag/svr/task_executor.py", line 442, in embedding
```
**Buggy code(/ragflow/rag/llm/embedding_model.py)**
```python
class HuggingFaceEmbed(Base):
def __init__(self, key, model_name, base_url=None):
if not model_name:
raise ValueError("Model name cannot be None")
self.key = key
self.model_name = model_name.split("___")[0]
self.base_url = base_url or "http://127.0.0.1:8080"
def encode(self, texts: list):
embeddings = []
for text in texts:
response = requests.post(...)
if response.status_code == 200:
try:
embedding = response.json()
embeddings.append(embedding[0])
# ❌ Early return
return np.array(embeddings), sum([num_tokens_from_string(text) for text in texts])
except Exception as _e:
log_exception(_e, response)
else:
raise Exception(...)
```
**Fixed Code(I just Rollback this function to the v0.19.0 version)**
```python
Class HuggingFaceEmbed(Base):
def __init__(self, key, model_name, base_url=None):
if not model_name:
raise ValueError("Model name cannot be None")
self.key = key
self.model_name = model_name.split("___")[0]
self.base_url = base_url or "http://127.0.0.1:8080"
def encode(self, texts: list):
embeddings = []
for text in texts:
response = requests.post(...)
if response.status_code == 200:
embedding = response.json()
embeddings.append(embedding[0]) # ✅ Only append, no return
else:
raise Exception(...)
return np.array(embeddings), sum([num_tokens_from_string(text) for text in texts]) # ✅ Return after processing all
```
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
This is a cherry-pick from #7781 as requested.
### 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?
- Simplify AzureChat constructor by passing base_url directly
- Clean up spacing and formatting in chat_model.py
- Remove redundant parentheses and improve code consistency
- #8423
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Fix auto-keyword and auto-question fail with qwq model. #8189
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Improve robustness of Jina, Nvidia, and SILICONFLOW embedding models by:
1. Adding try-catch blocks for JSON decode errors
2. Logging error details including response content
3. Raising exceptions with meaningful error messages
### Type of change
- [x] Refactoring
### 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?
Fix: local variable referenced before assignment. #6803
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
This PR introduces **primitive support for function calls**,
enabling the system to handle basic function call capabilities.
However, this feature is currently experimental and **not yet enabled
for general use**, as it is only supported by a subset of models,
namely, Qwen and OpenAI models.
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
### What problem does this PR solve?
https://github.com/infiniflow/ragflow/issues/6138
This PR is going to support vision llm for gpustack, modify url path
from `/v1-openai` to `/v1`
### Type of change
- [x] New Feature (non-breaking change which adds functionality)
# Dynamic Context Window Size for Ollama Chat
## Problem Statement
Previously, the Ollama chat implementation used a fixed context window
size of 32768 tokens. This caused two main issues:
1. Performance degradation due to unnecessarily large context windows
for small conversations
2. Potential business logic failures when using smaller fixed sizes
(e.g., 2048 tokens)
## Solution
Implemented a dynamic context window size calculation that:
1. Uses a base context size of 8192 tokens
2. Applies a 1.2x buffer ratio to the total token count
3. Adds multiples of 8192 tokens based on the buffered token count
4. Implements a smart context size update strategy
## Implementation Details
### Token Counting Logic
```python
def count_tokens(text):
"""Calculate token count for text"""
# Simple calculation: 1 token per ASCII character
# 2 tokens for non-ASCII characters (Chinese, Japanese, Korean, etc.)
total = 0
for char in text:
if ord(char) < 128: # ASCII characters
total += 1
else: # Non-ASCII characters
total += 2
return total
```
### Dynamic Context Calculation
```python
def _calculate_dynamic_ctx(self, history):
"""Calculate dynamic context window size"""
# Calculate total tokens for all messages
total_tokens = 0
for message in history:
content = message.get("content", "")
content_tokens = count_tokens(content)
role_tokens = 4 # Role marker token overhead
total_tokens += content_tokens + role_tokens
# Apply 1.2x buffer ratio
total_tokens_with_buffer = int(total_tokens * 1.2)
# Calculate context size in multiples of 8192
if total_tokens_with_buffer <= 8192:
ctx_size = 8192
else:
ctx_multiplier = (total_tokens_with_buffer // 8192) + 1
ctx_size = ctx_multiplier * 8192
return ctx_size
```
### Integration in Chat Method
```python
def chat(self, system, history, gen_conf):
if system:
history.insert(0, {"role": "system", "content": system})
if "max_tokens" in gen_conf:
del gen_conf["max_tokens"]
try:
# Calculate new context size
new_ctx_size = self._calculate_dynamic_ctx(history)
# Prepare options with context size
options = {
"num_ctx": new_ctx_size
}
# Add other generation options
if "temperature" in gen_conf:
options["temperature"] = gen_conf["temperature"]
if "max_tokens" in gen_conf:
options["num_predict"] = gen_conf["max_tokens"]
if "top_p" in gen_conf:
options["top_p"] = gen_conf["top_p"]
if "presence_penalty" in gen_conf:
options["presence_penalty"] = gen_conf["presence_penalty"]
if "frequency_penalty" in gen_conf:
options["frequency_penalty"] = gen_conf["frequency_penalty"]
# Make API call with dynamic context size
response = self.client.chat(
model=self.model_name,
messages=history,
options=options,
keep_alive=60
)
return response["message"]["content"].strip(), response.get("eval_count", 0) + response.get("prompt_eval_count", 0)
except Exception as e:
return "**ERROR**: " + str(e), 0
```
## Benefits
1. **Improved Performance**: Uses appropriate context windows based on
conversation length
2. **Better Resource Utilization**: Context window size scales with
content
3. **Maintained Compatibility**: Works with existing business logic
4. **Predictable Scaling**: Context growth in 8192-token increments
5. **Smart Updates**: Context size updates are optimized to reduce
unnecessary model reloads
## Future Considerations
1. Fine-tune buffer ratio based on usage patterns
2. Add monitoring for context window utilization
3. Consider language-specific token counting optimizations
4. Implement adaptive threshold based on conversation patterns
5. Add metrics for context size update frequency
---------
Co-authored-by: Kevin Hu <kevinhu.sh@gmail.com>
### What problem does this PR solve?
When using the online large model API knowledge base to extract
knowledge graphs, frequent Rate Limit Errors were triggered,
causing document parsing to fail. This commit fixes the issue by
optimizing API calls in the following way:
Added exponential backoff and jitter to the API call to reduce the
frequency of Rate Limit Errors.
### 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?
Fix the error where the Ollama embeddings interface returns a “500
Internal Server Error” when using models such as xiaobu-embedding-v2 for
embedding.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
### What problem does this PR solve?
Add vision LLM PDF parser
### 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?
vLLM provider with a reranking model does not work : as vLLM uses under
the hood the [CoHereRerank
provider](https://github.com/infiniflow/ragflow/blob/v0.17.0/rag/llm/__init__.py#L250)
with a `base_url`, if this URL [is not passed to the Cohere
client](https://github.com/infiniflow/ragflow/blob/v0.17.0/rag/llm/rerank_model.py#L379-L382)
any attempt will endup on the Cohere SaaS (sending your private api key
in the process) instead of your vLLM instance.
### 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):
