🐫 CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org
Python agentai-societiesartificial-intelligencecommunicative-aicooperative-aideep-learninglarge-language-modelsmulti-agent-systemsnatural-language-processing
feat: jina reranker refactor + api support (#2258)
JINO ROHIT885298b47cc9
diff --git a/camel/toolkits/jina_reranker_toolkit.py b/camel/toolkits/jina_reranker_toolkit.pyindex 8187ecaa8..02e443340 100644--- a/camel/toolkits/jina_reranker_toolkit.py+++ b/camel/toolkits/jina_reranker_toolkit.py@@ -11,7 +11,11 @@# See the License for the specific language governing permissions and# limitations under the License.# ========= Copyright 2023-2024 @ CAMEL-AI.org. All Rights Reserved. =========-from typing import List, Optional, Tuple+import json+import os+from typing import Any, Dict, List, Optional++import requestsfrom camel.toolkits import FunctionToolfrom camel.toolkits.base import BaseToolkit@@ -30,7 +34,9 @@ class JinaRerankerToolkit(BaseToolkit):def __init__(self,timeout: Optional[float] = None,+ model_name: Optional[str] = "jinaai/jina-reranker-m0",device: Optional[str] = None,+ use_api: bool = True,) -> None:r"""Initializes a new instance of the JinaRerankerToolkit class.@@ -38,31 +44,57 @@ class JinaRerankerToolkit(BaseToolkit):timeout (Optional[float]): The timeout value for API requestsin seconds. If None, no timeout is applied.(default: :obj:`None`)+ model_name (Optional[str]): The reranker model name. If None,+ will use the default model.+ (default: :obj:`None`)device (Optional[str]): Device to load the model on. If None,will use CUDA if available, otherwise CPU.+ Only effective when use_api=False.(default: :obj:`None`)+ use_api (bool): A flag to switch between local model and API.+ (default: :obj:`True`)"""- import torch- from transformers import AutoModelsuper().__init__(timeout=timeout)- self.model = AutoModel.from_pretrained(- 'jinaai/jina-reranker-m0',- torch_dtype="auto",- trust_remote_code=True,- )- DEVICE = (- device- if device is not None- else ("cuda" if torch.cuda.is_available() else "cpu")- )- self.model.to(DEVICE)- self.model.eval()+ self.use_api = use_api+ self.model_name = model_name++ if self.use_api:+ self.model = None+ self._api_key = os.environ.get("JINA_API_KEY", "None")+ if self._api_key == "None":+ raise ValueError(+ "Missing or empty required API keys in "+ "environment variables\n"+ "You can obtain the API key from https://jina.ai/reranker/"+ )+ self.url = 'https://api.jina.ai/v1/rerank'+ self.headers = {+ 'Content-Type': 'application/json',+ 'Accept': 'application/json',+ 'Authorization': f'Bearer {self._api_key}',+ }+ else:+ import torch+ from transformers import AutoModel++ self.model = AutoModel.from_pretrained(+ self.model_name,+ torch_dtype="auto",+ trust_remote_code=True,+ )+ self.device = (+ device+ if device is not None+ else ("cuda" if torch.cuda.is_available() else "cpu")+ )+ self.model.to(self.device)+ self.model.eval()def _sort_documents(self, documents: List[str], scores: List[float]- ) -> List[Tuple[str, float]]:+ ) -> List[Dict[str, object]]:r"""Sort documents by their scores in descending order.Args:@@ -70,7 +102,7 @@ class JinaRerankerToolkit(BaseToolkit):scores (List[float]): Corresponding scores for each document.Returns:- List[Tuple[str, float]]: Sorted list of (document, score) pairs.+ List[Dict[str, object]]: Sorted list of (document, score) pairs.Raises:ValueError: If documents and scores have different lengths.@@ -80,14 +112,45 @@ class JinaRerankerToolkit(BaseToolkit):doc_score_pairs = list(zip(documents, scores))doc_score_pairs.sort(key=lambda x: x[1], reverse=True)- return doc_score_pairs+ results = [+ {'document': {'text': doc}, 'relevance_score': score}+ for doc, score in doc_score_pairs+ ]++ return results++ def _call_jina_api(self, data: Dict[str, Any]) -> List[Dict[str, object]]:+ r"""Makes a call to the JINA API for reranking.++ Args:+ data (Dict[str]): The data to be passed into the api body.++ Returns:+ List[Dict[str, object]]: A list of dictionary containing+ the reranked documents and their relevance scores.+ """+ try:+ response = requests.post(+ self.url,+ headers=self.headers,+ data=json.dumps(data),+ timeout=self.timeout,+ )+ response.raise_for_status()+ results = [+ {key: value for key, value in _res.items() if key != 'index'}+ for _res in response.json()['results']+ ]+ return results+ except requests.exceptions.RequestException as e:+ raise RuntimeError(f"Failed to get response from Jina AI: {e}")def rerank_text_documents(self,query: str,documents: List[str],max_length: int = 1024,- ) -> List[Tuple[str, float]]:+ ) -> List[Dict[str, object]]:r"""Reranks text documents based on their relevance to a text query.Args:@@ -97,21 +160,34 @@ class JinaRerankerToolkit(BaseToolkit):(default: :obj:`1024`)Returns:- List[Tuple[str, float]]: A list of tuples containing+ List[Dict[str, object]]: A list of dictionary containingthe reranked documents and their relevance scores."""- import torch- if self.model is None:- raise ValueError(- "Model has not been initialized or failed to initialize."- )+ data = {+ 'model': self.model_name,+ 'query': query,+ 'top_n': len(documents),+ 'documents': documents,+ 'return_documents': True,+ }- with torch.inference_mode():- text_pairs = [[query, doc] for doc in documents]- scores = self.model.compute_score(- text_pairs, max_length=max_length, doc_type="text"- )+ if self.use_api:+ return self._call_jina_api(data)++ else:+ import torch++ if self.model is None:+ raise ValueError(+ "Model has not been initialized or failed to initialize."+ )++ with torch.inference_mode():+ text_pairs = [[query, doc] for doc in documents]+ scores = self.model.compute_score(+ text_pairs, max_length=max_length, doc_type="text"+ )return self._sort_documents(documents, scores)@@ -120,7 +196,7 @@ class JinaRerankerToolkit(BaseToolkit):query: str,documents: List[str],max_length: int = 2048,- ) -> List[Tuple[str, float]]:+ ) -> List[Dict[str, object]]:r"""Reranks image documents based on their relevance to a text query.Args:@@ -130,21 +206,33 @@ class JinaRerankerToolkit(BaseToolkit):(default: :obj:`2048`)Returns:- List[Tuple[str, float]]: A list of tuples containing+ List[Dict[str, object]]: A list of dictionary containingthe reranked image URLs/paths and their relevance scores."""- import torch-- if self.model is None:- raise ValueError(- "Model has not been initialized or failed to initialize."- )-- with torch.inference_mode():- image_pairs = [[query, doc] for doc in documents]- scores = self.model.compute_score(- image_pairs, max_length=max_length, doc_type="image"- )+ data = {+ 'model': self.model_name,+ 'query': query,+ 'top_n': len(documents),+ 'documents': documents,+ 'return_documents': True,+ }++ if self.use_api:+ return self._call_jina_api(data)++ else:+ import torch++ if self.model is None:+ raise ValueError(+ "Model has not been initialized or failed to initialize."+ )++ with torch.inference_mode():+ image_pairs = [[query, doc] for doc in documents]+ scores = self.model.compute_score(+ image_pairs, max_length=max_length, doc_type="image"+ )return self._sort_documents(documents, scores)@@ -153,7 +241,7 @@ class JinaRerankerToolkit(BaseToolkit):image_query: str,documents: List[str],max_length: int = 2048,- ) -> List[Tuple[str, float]]:+ ) -> List[Dict[str, object]]:r"""Reranks text documents based on their relevance to an image query.Args:@@ -163,30 +251,45 @@ class JinaRerankerToolkit(BaseToolkit):(default: :obj:`2048`)Returns:- List[Tuple[str, float]]: A list of tuples containing+ List[Dict[str, object]]: A list of dictionary containingthe reranked documents and their relevance scores."""- import torch-- if self.model is None:- raise ValueError("Model has not been initialized.")- with torch.inference_mode():- image_pairs = [[image_query, doc] for doc in documents]- scores = self.model.compute_score(- image_pairs,- max_length=max_length,- query_type="image",- doc_type="text",- )-- return self._sort_documents(documents, scores)+ data = {+ 'model': self.model_name,+ 'query': image_query,+ 'top_n': len(documents),+ 'documents': documents,+ 'return_documents': True,+ }++ if self.use_api:+ return self._call_jina_api(data)++ else:+ import torch++ if self.model is None:+ raise ValueError(+ "Model has not been initialized or failed to initialize."+ )++ with torch.inference_mode():+ image_pairs = [[image_query, doc] for doc in documents]+ scores = self.model.compute_score(+ image_pairs,+ max_length=max_length,+ query_type="image",+ doc_type="text",+ )++ return self._sort_documents(documents, scores)def image_query_image_documents(self,image_query: str,documents: List[str],max_length: int = 2048,- ) -> List[Tuple[str, float]]:+ ) -> List[Dict[str, object]]:r"""Reranks image documents based on their relevance to an image query.Args:@@ -196,24 +299,38 @@ class JinaRerankerToolkit(BaseToolkit):(default: :obj:`2048`)Returns:- List[Tuple[str, float]]: A list of tuples containing+ List[Dict[str, object]]: A list of dictionary containingthe reranked image URLs/paths and their relevance scores."""- import torch-- if self.model is None:- raise ValueError("Model has not been initialized.")-- with torch.inference_mode():- image_pairs = [[image_query, doc] for doc in documents]- scores = self.model.compute_score(- image_pairs,- max_length=max_length,- query_type="image",- doc_type="image",- )-- return self._sort_documents(documents, scores)+ data = {+ 'model': self.model_name,+ 'query': image_query,+ 'top_n': len(documents),+ 'documents': documents,+ 'return_documents': True,+ }++ if self.use_api:+ return self._call_jina_api(data)++ else:+ import torch++ if self.model is None:+ raise ValueError(+ "Model has not been initialized or failed to initialize."+ )++ with torch.inference_mode():+ image_pairs = [[image_query, doc] for doc in documents]+ scores = self.model.compute_score(+ image_pairs,+ max_length=max_length,+ query_type="image",+ doc_type="image",+ )++ return self._sort_documents(documents, scores)def get_tools(self) -> List[FunctionTool]:r"""Returns a list of FunctionTool objects representing thediff --git a/examples/toolkits/jina_reranker_toolkit.py b/examples/toolkits/jina_reranker_toolkit.pyindex fefc4ccf7..18db17ea7 100644--- a/examples/toolkits/jina_reranker_toolkit.py+++ b/examples/toolkits/jina_reranker_toolkit.py@@ -22,7 +22,7 @@ model = ModelFactory.create(model_type=ModelType.DEFAULT,)-reranker_toolkit = JinaRerankerToolkit(device="cpu")+reranker_toolkit = JinaRerankerToolkit(device="cpu", use_api=False)reranker_tool = reranker_toolkit.get_tools()agent = ChatAgent(model=model, tools=reranker_tool)diff --git a/test/toolkits/test_jina_reranker_toolkit.py b/test/toolkits/test_jina_reranker_toolkit.pyindex cf9549ba1..bcc4c7a6e 100644--- a/test/toolkits/test_jina_reranker_toolkit.py+++ b/test/toolkits/test_jina_reranker_toolkit.py@@ -32,7 +32,7 @@ def reranker_toolkit(mock_model):with patch('transformers.AutoModel.from_pretrained', return_value=mock_model):- toolkit = JinaRerankerToolkit()+ toolkit = JinaRerankerToolkit(use_api=False)# Replace the actual model with our mocktoolkit.model = mock_modelreturn toolkit@@ -44,7 +44,7 @@ def test_init_with_default_params():'transformers.AutoModel.from_pretrained') as mock_from_pretrained:with patch('torch.cuda.is_available', return_value=False):- JinaRerankerToolkit()+ JinaRerankerToolkit(use_api=False)mock_from_pretrained.assert_called_once_with('jinaai/jina-reranker-m0',torch_dtype="auto",@@ -61,7 +61,7 @@ def test_init_with_custom_device():model_mock = MagicMock()mock_from_pretrained.return_value = model_mock- JinaRerankerToolkit(device="cpu")+ JinaRerankerToolkit(use_api=False, device="cpu")model_mock.to.assert_called_once_with("cpu")model_mock.eval.assert_called_once()@@ -72,14 +72,18 @@ def test_init_with_custom_device():)def test_sort_documents():r"""Test the _sort_documents method."""- toolkit = JinaRerankerToolkit()+ toolkit = JinaRerankerToolkit(use_api=False)documents = ["doc1", "doc2", "doc3"]scores = [0.5, 0.9, 0.1]result = toolkit._sort_documents(documents, scores)# Should be sorted by score in descending order- assert result == [("doc2", 0.9), ("doc1", 0.5), ("doc3", 0.1)]+ assert result == [+ {"document": "doc2", "score": 0.9},+ {"document": "doc1", "score": 0.5},+ {"document": "doc3", "score": 0.1},+ ]@pytest.mark.skip(@@ -87,7 +91,7 @@ def test_sort_documents():)def test_sort_documents_with_mismatched_lengths():r"""Test _sort_documents with mismatched document and score lists."""- toolkit = JinaRerankerToolkit()+ toolkit = JinaRerankerToolkit(use_api=False)documents = ["doc1", "doc2"]scores = [0.5, 0.9, 0.1]@@ -113,7 +117,11 @@ def test_rerank_text_documents(reranker_toolkit):)# Check that the result is sorted correctly- assert result == [("doc3", 0.9), ("doc1", 0.7), ("doc2", 0.3)]+ assert result == [+ {'document': {'text': 'doc3'}, 'relevance_score': 0.9},+ {'document': {'text': 'doc1'}, 'relevance_score': 0.7},+ {'document': {'text': 'doc2'}, 'relevance_score': 0.3},+ ]def test_rerank_text_documents_with_custom_max_length(reranker_toolkit):@@ -132,7 +140,10 @@ def test_rerank_text_documents_with_custom_max_length(reranker_toolkit):[[query, doc] for doc in documents], max_length=512, doc_type="text")- assert result == [("doc2", 0.8), ("doc1", 0.6)]+ assert result == [+ {'document': {'text': 'doc2'}, 'relevance_score': 0.8},+ {'document': {'text': 'doc1'}, 'relevance_score': 0.6},+ ]@pytest.mark.skip(@@ -140,7 +151,7 @@ def test_rerank_text_documents_with_custom_max_length(reranker_toolkit):)def test_rerank_text_documents_model_not_initialized():r"""Test reranking text documents when model is not initialized."""- toolkit = JinaRerankerToolkit()+ toolkit = JinaRerankerToolkit(use_api=False)toolkit.model = Nonewith pytest.raises(@@ -164,7 +175,10 @@ def test_rerank_image_documents(reranker_toolkit):[[query, doc] for doc in documents], max_length=2048, doc_type="image")- assert result == [("image2.jpg", 0.8), ("image1.jpg", 0.4)]+ assert result == [+ {'document': {'text': 'image2.jpg'}, 'relevance_score': 0.8},+ {'document': {'text': 'image1.jpg'}, 'relevance_score': 0.4},+ ]def test_rerank_image_documents_custom_max_length(reranker_toolkit):@@ -183,7 +197,10 @@ def test_rerank_image_documents_custom_max_length(reranker_toolkit):[[query, doc] for doc in documents], max_length=1024, doc_type="image")- assert result == [("image2.jpg", 0.8), ("image1.jpg", 0.4)]+ assert result == [+ {'document': {'text': 'image2.jpg'}, 'relevance_score': 0.8},+ {'document': {'text': 'image1.jpg'}, 'relevance_score': 0.4},+ ]def test_image_query_text_documents(reranker_toolkit):@@ -205,7 +222,10 @@ def test_image_query_text_documents(reranker_toolkit):doc_type="text",)- assert result == [("doc2", 0.7), ("doc1", 0.3)]+ assert result == [+ {'document': {'text': 'doc2'}, 'relevance_score': 0.7},+ {'document': {'text': 'doc1'}, 'relevance_score': 0.3},+ ]def test_image_query_image_documents(reranker_toolkit):@@ -227,7 +247,10 @@ def test_image_query_image_documents(reranker_toolkit):doc_type="image",)- assert result == [("image1.jpg", 0.6), ("image2.jpg", 0.2)]+ assert result == [+ {'document': {'text': 'image1.jpg'}, 'relevance_score': 0.6},+ {'document': {'text': 'image2.jpg'}, 'relevance_score': 0.2},+ ]def test_with_inference_mode(reranker_toolkit):