agno 2.1.2__py3-none-any.whl → 2.3.13__py3-none-any.whl

agno/knowledge/chunking/strategy.py

@@ -1,6 +1,6 @@
 from abc import ABC, abstractmethod
 from enum import Enum
-from typing import List
+from typing import List, Optional
 
 from agno.knowledge.document.base import Document
 
@@ -60,7 +60,13 @@ class ChunkingStrategyFactory:
     """Factory for creating chunking strategy instances."""
 
     @classmethod
-    def create_strategy(cls, strategy_type: ChunkingStrategyType, **kwargs) -> ChunkingStrategy:
+    def create_strategy(
+        cls,
+        strategy_type: ChunkingStrategyType,
+        chunk_size: Optional[int] = None,
+        overlap: Optional[int] = None,
+        **kwargs,
+    ) -> ChunkingStrategy:
         """Create an instance of the chunking strategy with the given parameters."""
         strategy_map = {
             ChunkingStrategyType.AGENTIC_CHUNKER: cls._create_agentic_chunking,
@@ -71,51 +77,89 @@ class ChunkingStrategyFactory:
             ChunkingStrategyType.ROW_CHUNKER: cls._create_row_chunking,
             ChunkingStrategyType.MARKDOWN_CHUNKER: cls._create_markdown_chunking,
         }
-        return strategy_map[strategy_type](**kwargs)
+        return strategy_map[strategy_type](chunk_size=chunk_size, overlap=overlap, **kwargs)
 
     @classmethod
-    def _create_agentic_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_agentic_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.agentic import AgenticChunking
 
-        # Map chunk_size to max_chunk_size for AgenticChunking
-        if "chunk_size" in kwargs and "max_chunk_size" not in kwargs:
-            kwargs["max_chunk_size"] = kwargs.pop("chunk_size")
+        # AgenticChunking accepts max_chunk_size (not chunk_size) and no overlap
+        if chunk_size is not None:
+            kwargs["max_chunk_size"] = chunk_size
+        # Remove overlap since AgenticChunking doesn't support it
         return AgenticChunking(**kwargs)
 
     @classmethod
-    def _create_document_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_document_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.document import DocumentChunking
 
+        # DocumentChunking accepts both chunk_size and overlap
+        if chunk_size is not None:
+            kwargs["chunk_size"] = chunk_size
+        if overlap is not None:
+            kwargs["overlap"] = overlap
         return DocumentChunking(**kwargs)
 
     @classmethod
-    def _create_recursive_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_recursive_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.recursive import RecursiveChunking
 
+        # RecursiveChunking accepts both chunk_size and overlap
+        if chunk_size is not None:
+            kwargs["chunk_size"] = chunk_size
+        if overlap is not None:
+            kwargs["overlap"] = overlap
         return RecursiveChunking(**kwargs)
 
     @classmethod
-    def _create_semantic_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_semantic_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.semantic import SemanticChunking
 
+        # SemanticChunking accepts chunk_size but not overlap
+        if chunk_size is not None:
+            kwargs["chunk_size"] = chunk_size
+        # Remove overlap since SemanticChunking doesn't support it
         return SemanticChunking(**kwargs)
 
     @classmethod
-    def _create_fixed_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_fixed_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.fixed import FixedSizeChunking
 
+        # FixedSizeChunking accepts both chunk_size and overlap
+        if chunk_size is not None:
+            kwargs["chunk_size"] = chunk_size
+        if overlap is not None:
+            kwargs["overlap"] = overlap
         return FixedSizeChunking(**kwargs)
 
     @classmethod
-    def _create_row_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_row_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.row import RowChunking
 
-        # Remove chunk_size if present since RowChunking doesn't use it
-        kwargs.pop("chunk_size", None)
+        # RowChunking doesn't accept chunk_size or overlap, only skip_header and clean_rows
         return RowChunking(**kwargs)
 
     @classmethod
-    def _create_markdown_chunking(cls, **kwargs) -> ChunkingStrategy:
+    def _create_markdown_chunking(
+        cls, chunk_size: Optional[int] = None, overlap: Optional[int] = None, **kwargs
+    ) -> ChunkingStrategy:
         from agno.knowledge.chunking.markdown import MarkdownChunking
 
+        # MarkdownChunking accepts both chunk_size and overlap
+        if chunk_size is not None:
+            kwargs["chunk_size"] = chunk_size
+        if overlap is not None:
+            kwargs["overlap"] = overlap
         return MarkdownChunking(**kwargs)

agno/knowledge/embedder/fastembed.py

@@ -23,7 +23,7 @@ class FastEmbedEmbedder(Embedder):
     """Using BAAI/bge-small-en-v1.5 model, more models available: https://qdrant.github.io/fastembed/examples/Supported_Models/"""
 
     id: str = "BAAI/bge-small-en-v1.5"
-    dimensions: int = 384
+    dimensions: Optional[int] = 384
 
     def get_embedding(self, text: str) -> List[float]:
         model = TextEmbedding(model_name=self.id)

agno/knowledge/embedder/nebius.py

@@ -10,4 +10,4 @@ class NebiusEmbedder(OpenAIEmbedder):
     id: str = "BAAI/bge-en-icl"
     dimensions: int = 1024
     api_key: Optional[str] = getenv("NEBIUS_API_KEY")
-    base_url: str = "https://api.studio.nebius.com/v1/"
+    base_url: str = "https://api.tokenfactory.nebius.com/v1/"

agno/knowledge/embedder/ollama.py

@@ -85,6 +85,10 @@ class OllamaEmbedder(Embedder):
         if self.options is not None:
             kwargs["options"] = self.options
 
+        # Add dimensions parameter for models that support it
+        if self.dimensions is not None:
+            kwargs["dimensions"] = self.dimensions
+
         response = self.client.embed(input=text, model=self.id, **kwargs)
         if response and "embeddings" in response:
             embeddings = response["embeddings"]
@@ -117,6 +121,10 @@ class OllamaEmbedder(Embedder):
         if self.options is not None:
             kwargs["options"] = self.options
 
+        # Add dimensions parameter for models that support it
+        if self.dimensions is not None:
+            kwargs["dimensions"] = self.dimensions
+
         response = await self.aclient.embed(input=text, model=self.id, **kwargs)
         if response and "embeddings" in response:
             embeddings = response["embeddings"]

agno/knowledge/embedder/openai.py

@@ -4,7 +4,7 @@ from typing import Any, Dict, List, Optional, Tuple
 from typing_extensions import Literal
 
 from agno.knowledge.embedder.base import Embedder
-from agno.utils.log import logger
+from agno.utils.log import log_info, log_warning
 
 try:
     from openai import AsyncOpenAI
@@ -82,7 +82,7 @@ class OpenAIEmbedder(Embedder):
             response: CreateEmbeddingResponse = self.response(text=text)
             return response.data[0].embedding
         except Exception as e:
-            logger.warning(e)
+            log_warning(e)
             return []
 
     def get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict]]:
@@ -95,7 +95,7 @@ class OpenAIEmbedder(Embedder):
                 return embedding, usage.model_dump()
             return embedding, None
         except Exception as e:
-            logger.warning(e)
+            log_warning(e)
             return [], None
 
     async def async_get_embedding(self, text: str) -> List[float]:
@@ -115,7 +115,7 @@ class OpenAIEmbedder(Embedder):
             response: CreateEmbeddingResponse = await self.aclient.embeddings.create(**req)
             return response.data[0].embedding
         except Exception as e:
-            logger.warning(e)
+            log_warning(e)
             return []
 
     async def async_get_embedding_and_usage(self, text: str):
@@ -137,7 +137,7 @@ class OpenAIEmbedder(Embedder):
             usage = response.usage
             return embedding, usage.model_dump() if usage else None
         except Exception as e:
-            logger.warning(e)
+            log_warning(f"Error getting embedding: {e}")
             return [], None
 
     async def async_get_embeddings_batch_and_usage(
@@ -154,7 +154,7 @@ class OpenAIEmbedder(Embedder):
         """
         all_embeddings = []
         all_usage = []
-        logger.info(f"Getting embeddings and usage for {len(texts)} texts in batches of {self.batch_size} (async)")
+        log_info(f"Getting embeddings and usage for {len(texts)} texts in batches of {self.batch_size} (async)")
 
         for i in range(0, len(texts), self.batch_size):
             batch_texts = texts[i : i + self.batch_size]
@@ -180,7 +180,7 @@ class OpenAIEmbedder(Embedder):
                 usage_dict = response.usage.model_dump() if response.usage else None
                 all_usage.extend([usage_dict] * len(batch_embeddings))
             except Exception as e:
-                logger.warning(f"Error in async batch embedding: {e}")
+                log_warning(f"Error in async batch embedding: {e}")
                 # Fallback to individual calls for this batch
                 for text in batch_texts:
                     try:
@@ -188,7 +188,7 @@ class OpenAIEmbedder(Embedder):
                         all_embeddings.append(embedding)
                         all_usage.append(usage)
                     except Exception as e2:
-                        logger.warning(f"Error in individual async embedding fallback: {e2}")
+                        log_warning(f"Error in individual async embedding fallback: {e2}")
                         all_embeddings.append([])
                         all_usage.append(None)
 

agno/knowledge/embedder/sentence_transformer.py

@@ -25,10 +25,14 @@ class SentenceTransformerEmbedder(Embedder):
     prompt: Optional[str] = None
     normalize_embeddings: bool = False
 
-    def get_embedding(self, text: Union[str, List[str]]) -> List[float]:
-        if not self.sentence_transformer_client:
+    def __post_init__(self):
+        # Initialize the SentenceTransformer model eagerly to avoid race conditions in async contexts
+        if self.sentence_transformer_client is None:
             self.sentence_transformer_client = SentenceTransformer(model_name_or_path=self.id)
 
+    def get_embedding(self, text: Union[str, List[str]]) -> List[float]:
+        if self.sentence_transformer_client is None:
+            raise RuntimeError("SentenceTransformer model not initialized")
         model = self.sentence_transformer_client
         embedding = model.encode(text, prompt=self.prompt, normalize_embeddings=self.normalize_embeddings)
         try:

agno/knowledge/embedder/vllm.py

@@ -0,0 +1,262 @@
+import asyncio
+from dataclasses import dataclass
+from os import getenv
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+
+from agno.knowledge.embedder.base import Embedder
+from agno.utils.log import logger
+
+try:
+    from vllm import LLM  # type: ignore
+    from vllm.outputs import EmbeddingRequestOutput  # type: ignore
+except ImportError:
+    raise ImportError("`vllm` not installed. Please install using `pip install vllm`.")
+
+if TYPE_CHECKING:
+    from openai import AsyncOpenAI
+    from openai import OpenAI as OpenAIClient
+    from openai.types.create_embedding_response import CreateEmbeddingResponse
+
+
+@dataclass
+class VLLMEmbedder(Embedder):
+    """
+    VLLM Embedder supporting both local and remote deployment modes.
+
+    Local Mode (default):
+        - Loads model locally and runs inference on your GPU/CPU
+        - No API key required
+        - Example: VLLMEmbedder(id="intfloat/e5-mistral-7b-instruct")
+
+    Remote Mode:
+        - Connects to a remote vLLM server via OpenAI-compatible API
+        - Uses OpenAI SDK to communicate with vLLM's OpenAI-compatible endpoint
+        - Requires base_url and optionally api_key
+        - Example: VLLMEmbedder(base_url="http://localhost:8000/v1", api_key="your-key")
+        - Ref: https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html
+    """
+
+    id: str = "sentence-transformers/all-MiniLM-L6-v2"
+    dimensions: int = 4096
+    # Local mode parameters
+    enforce_eager: bool = True
+    vllm_kwargs: Optional[Dict[str, Any]] = None
+    vllm_client: Optional[LLM] = None
+    # Remote mode parameters
+    api_key: Optional[str] = getenv("VLLM_API_KEY")
+    base_url: Optional[str] = None
+    request_params: Optional[Dict[str, Any]] = None
+    client_params: Optional[Dict[str, Any]] = None
+    remote_client: Optional["OpenAIClient"] = None  # OpenAI-compatible client for vLLM server
+    async_remote_client: Optional["AsyncOpenAI"] = None  # Async OpenAI-compatible client for vLLM server
+
+    @property
+    def is_remote(self) -> bool:
+        """Determine if we should use remote mode."""
+        return self.base_url is not None
+
+    def _get_vllm_client(self) -> LLM:
+        """Get local VLLM client."""
+        if self.vllm_client:
+            return self.vllm_client
+
+        _vllm_params: Dict[str, Any] = {
+            "model": self.id,
+            "task": "embed",
+            "enforce_eager": self.enforce_eager,
+        }
+        if self.vllm_kwargs:
+            _vllm_params.update(self.vllm_kwargs)
+        self.vllm_client = LLM(**_vllm_params)
+        return self.vllm_client
+
+    def _get_remote_client(self) -> "OpenAIClient":
+        """Get OpenAI-compatible client for remote vLLM server."""
+        if self.remote_client:
+            return self.remote_client
+
+        try:
+            from openai import OpenAI as OpenAIClient
+        except ImportError:
+            raise ImportError("`openai` package required for remote vLLM mode. ")
+
+        _client_params: Dict[str, Any] = {
+            "api_key": self.api_key or "EMPTY",  # VLLM can run without API key
+            "base_url": self.base_url,
+        }
+        if self.client_params:
+            _client_params.update(self.client_params)
+        self.remote_client = OpenAIClient(**_client_params)
+        return self.remote_client
+
+    def _get_async_remote_client(self) -> "AsyncOpenAI":
+        """Get async OpenAI-compatible client for remote vLLM server."""
+        if self.async_remote_client:
+            return self.async_remote_client
+
+        try:
+            from openai import AsyncOpenAI
+        except ImportError:
+            raise ImportError("`openai` package required for remote vLLM mode. ")
+
+        _client_params: Dict[str, Any] = {
+            "api_key": self.api_key or "EMPTY",
+            "base_url": self.base_url,
+        }
+        if self.client_params:
+            _client_params.update(self.client_params)
+        self.async_remote_client = AsyncOpenAI(**_client_params)
+        return self.async_remote_client
+
+    def _create_embedding_local(self, text: str) -> Optional[EmbeddingRequestOutput]:
+        """Create embedding using local VLLM."""
+        try:
+            outputs = self._get_vllm_client().embed([text])
+            return outputs[0] if outputs else None
+        except Exception as e:
+            logger.warning(f"Error creating local embedding: {e}")
+            return None
+
+    def _create_embedding_remote(self, text: str) -> "CreateEmbeddingResponse":
+        """Create embedding using remote vLLM server."""
+        _request_params: Dict[str, Any] = {
+            "input": text,
+            "model": self.id,
+        }
+        if self.request_params:
+            _request_params.update(self.request_params)
+        return self._get_remote_client().embeddings.create(**_request_params)
+
+    def get_embedding(self, text: str) -> List[float]:
+        try:
+            if self.is_remote:
+                # Remote mode: OpenAI-compatible API
+                response: "CreateEmbeddingResponse" = self._create_embedding_remote(text=text)
+                return response.data[0].embedding
+            else:
+                # Local mode: Direct VLLM
+                output = self._create_embedding_local(text=text)
+                if output and hasattr(output, "outputs") and hasattr(output.outputs, "embedding"):
+                    embedding = output.outputs.embedding
+                    if len(embedding) != self.dimensions:
+                        logger.warning(f"Expected embedding dimension {self.dimensions}, but got {len(embedding)}")
+                    return embedding
+                return []
+        except Exception as e:
+            logger.warning(f"Error extracting embedding: {e}")
+            return []
+
+    def get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict]]:
+        if self.is_remote:
+            try:
+                response: "CreateEmbeddingResponse" = self._create_embedding_remote(text=text)
+                embedding = response.data[0].embedding
+                usage = response.usage
+                if usage:
+                    return embedding, usage.model_dump()
+                return embedding, None
+            except Exception as e:
+                logger.warning(f"Error in remote embedding: {e}")
+                return [], None
+        else:
+            embedding = self.get_embedding(text=text)
+            # Local VLLM doesn't provide usage information
+            return embedding, None
+
+    async def async_get_embedding(self, text: str) -> List[float]:
+        """Async version of get_embedding using thread executor for local mode."""
+        if self.is_remote:
+            # Remote mode: async client for vLLM server
+            try:
+                req: Dict[str, Any] = {
+                    "input": text,
+                    "model": self.id,
+                }
+                if self.request_params:
+                    req.update(self.request_params)
+                response: "CreateEmbeddingResponse" = await self._get_async_remote_client().embeddings.create(**req)
+                return response.data[0].embedding
+            except Exception as e:
+                logger.warning(f"Error in async remote embedding: {e}")
+                return []
+        else:
+            # Local mode: use thread executor for CPU-bound operations
+            loop = asyncio.get_event_loop()
+            return await loop.run_in_executor(None, self.get_embedding, text)
+
+    async def async_get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict]]:
+        """Async version of get_embedding_and_usage using thread executor for local mode."""
+        if self.is_remote:
+            try:
+                req: Dict[str, Any] = {
+                    "input": text,
+                    "model": self.id,
+                }
+                if self.request_params:
+                    req.update(self.request_params)
+                response: "CreateEmbeddingResponse" = await self._get_async_remote_client().embeddings.create(**req)
+                embedding = response.data[0].embedding
+                usage = response.usage
+                return embedding, usage.model_dump() if usage else None
+            except Exception as e:
+                logger.warning(f"Error in async remote embedding: {e}")
+                return [], None
+        else:
+            # Local mode: use thread executor for CPU-bound operations
+            try:
+                loop = asyncio.get_event_loop()
+                return await loop.run_in_executor(None, self.get_embedding_and_usage, text)
+            except Exception as e:
+                logger.warning(f"Error in async local embedding: {e}")
+                return [], None
+
+    async def async_get_embeddings_batch_and_usage(
+        self, texts: List[str]
+    ) -> Tuple[List[List[float]], List[Optional[Dict]]]:
+        """
+        Get embeddings and usage for multiple texts in batches (async version).
+
+        Args:
+            texts: List of text strings to embed
+
+        Returns:
+            Tuple of (List of embedding vectors, List of usage dictionaries)
+        """
+        all_embeddings = []
+        all_usage = []
+        logger.info(f"Getting embeddings for {len(texts)} texts in batches of {self.batch_size} (async)")
+
+        for i in range(0, len(texts), self.batch_size):
+            batch_texts = texts[i : i + self.batch_size]
+
+            try:
+                if self.is_remote:
+                    # Remote mode: use batch API
+                    req: Dict[str, Any] = {
+                        "input": batch_texts,
+                        "model": self.id,
+                    }
+                    if self.request_params:
+                        req.update(self.request_params)
+                    response: "CreateEmbeddingResponse" = await self._get_async_remote_client().embeddings.create(**req)
+                    batch_embeddings = [data.embedding for data in response.data]
+                    all_embeddings.extend(batch_embeddings)
+
+                    # For each embedding in the batch, add the same usage information
+                    usage_dict = response.usage.model_dump() if response.usage else None
+                    all_usage.extend([usage_dict] * len(batch_embeddings))
+                else:
+                    # Local mode: process individually using thread executor
+                    for text in batch_texts:
+                        embedding, usage = await self.async_get_embedding_and_usage(text)
+                        all_embeddings.append(embedding)
+                        all_usage.append(usage)
+
+            except Exception as e:
+                logger.warning(f"Error in async batch embedding: {e}")
+                # Fallback: add empty results for failed batch
+                for _ in batch_texts:
+                    all_embeddings.append([])
+                    all_usage.append(None)
+
+        return all_embeddings, all_usage