agno 2.0.11__py3-none-any.whl → 2.1.1__py3-none-any.whl

agno/knowledge/embedder/cohere.py

@@ -1,8 +1,9 @@
+import time
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Tuple, Union
 
 from agno.knowledge.embedder.base import Embedder
-from agno.utils.log import logger
+from agno.utils.log import log_debug, log_error, log_info, log_warning
 
 try:
     from cohere import AsyncClient as AsyncCohereClient
@@ -22,6 +23,7 @@ class CohereEmbedder(Embedder):
     client_params: Optional[Dict[str, Any]] = None
     cohere_client: Optional[CohereClient] = None
     async_client: Optional[AsyncCohereClient] = None
+    exponential_backoff: bool = False  # Enable exponential backoff on rate limits
 
     @property
     def client(self) -> CohereClient:
@@ -61,6 +63,111 @@ class CohereEmbedder(Embedder):
             request_params.update(self.request_params)
         return self.client.embed(texts=[text], **request_params)
 
+    def _get_batch_request_params(self) -> Dict[str, Any]:
+        """Get request parameters for batch embedding calls."""
+        request_params: Dict[str, Any] = {}
+
+        if self.id:
+            request_params["model"] = self.id
+        if self.input_type:
+            request_params["input_type"] = self.input_type
+        if self.embedding_types:
+            request_params["embedding_types"] = self.embedding_types
+        if self.request_params:
+            request_params.update(self.request_params)
+
+        return request_params
+
+    def _is_rate_limit_error(self, error: Exception) -> bool:
+        """Check if the error is a rate limiting error."""
+        if hasattr(error, "status_code") and error.status_code == 429:
+            return True
+        error_str = str(error).lower()
+        return any(
+            phrase in error_str
+            for phrase in ["rate limit", "too many requests", "429", "trial key", "api calls / minute"]
+        )
+
+    def _exponential_backoff_sleep(self, attempt: int, base_delay: float = 1.0) -> None:
+        """Sleep with exponential backoff."""
+        delay = base_delay * (2**attempt) + (time.time() % 1)  # Add jitter
+        log_debug(f"Rate limited, waiting {delay:.2f} seconds before retry (attempt {attempt + 1})")
+        time.sleep(delay)
+
+    async def _async_rate_limit_backoff_sleep(self, attempt: int) -> None:
+        """Async version of rate-limit-aware backoff for APIs with per-minute limits."""
+        import asyncio
+
+        # For 40 req/min APIs like Cohere Trial, we need longer waits
+        if attempt == 0:
+            delay = 15.0  # Wait 15 seconds (1/4 of minute window)
+        elif attempt == 1:
+            delay = 30.0  # Wait 30 seconds (1/2 of minute window)
+        else:
+            delay = 60.0  # Wait full minute for window reset
+
+        # Add small jitter
+        delay += time.time() % 3
+
+        log_debug(
+            f"Async rate limit backoff, waiting {delay:.1f} seconds for rate limit window reset (attempt {attempt + 1})"
+        )
+        await asyncio.sleep(delay)
+
+    async def _async_batch_with_retry(
+        self, texts: List[str], max_retries: int = 3
+    ) -> Tuple[List[List[float]], List[Optional[Dict]]]:
+        """Execute async batch embedding with rate-limit-aware backoff for rate limiting."""
+
+        log_debug(f"Starting async batch retry for {len(texts)} texts with max_retries={max_retries}")
+
+        for attempt in range(max_retries + 1):
+            try:
+                request_params = self._get_batch_request_params()
+                response: Union[
+                    EmbeddingsFloatsEmbedResponse, EmbeddingsByTypeEmbedResponse
+                ] = await self.aclient.embed(texts=texts, **request_params)
+
+                # Extract embeddings from response
+                if isinstance(response, EmbeddingsFloatsEmbedResponse):
+                    batch_embeddings = response.embeddings
+                elif isinstance(response, EmbeddingsByTypeEmbedResponse):
+                    batch_embeddings = response.embeddings.float_ if response.embeddings.float_ else []
+                else:
+                    log_warning("No embeddings found in response")
+                    batch_embeddings = []
+
+                # Extract usage information
+                usage = response.meta.billed_units if response.meta else None
+                usage_dict = usage.model_dump() if usage else None
+                all_usage = [usage_dict] * len(batch_embeddings)
+
+                log_debug(f"Async batch embedding succeeded on attempt {attempt + 1}")
+                return batch_embeddings, all_usage
+
+            except Exception as e:
+                if self._is_rate_limit_error(e):
+                    if not self.exponential_backoff:
+                        log_warning(
+                            "Rate limit detected. To enable automatic backoff retry, set enable_backoff=True when creating the embedder."
+                        )
+                        raise e
+
+                    log_info(f"Async rate limit detected on attempt {attempt + 1}")
+                    if attempt < max_retries:
+                        await self._async_rate_limit_backoff_sleep(attempt)
+                        continue
+                    else:
+                        log_warning(f"Async max retries ({max_retries}) reached for rate limiting")
+                        raise e
+                else:
+                    log_debug(f"Async non-rate-limit error on attempt {attempt + 1}: {e}")
+                    raise e
+
+        # This should never be reached, but just in case
+        log_error("Could not create embeddings. End of retry loop reached.")
+        return [], []
+
     def get_embedding(self, text: str) -> List[float]:
         response: Union[EmbeddingsFloatsEmbedResponse, EmbeddingsByTypeEmbedResponse] = self.response(text=text)
         try:
@@ -69,10 +176,10 @@ class CohereEmbedder(Embedder):
             elif isinstance(response, EmbeddingsByTypeEmbedResponse):
                 return response.embeddings.float_[0] if response.embeddings.float_ else []
             else:
-                logger.warning("No embeddings found")
+                log_warning("No embeddings found")
                 return []
         except Exception as e:
-            logger.warning(e)
+            log_warning(e)
             return []
 
     def get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict[str, Any]]]:
@@ -110,10 +217,10 @@ class CohereEmbedder(Embedder):
             elif isinstance(response, EmbeddingsByTypeEmbedResponse):
                 return response.embeddings.float_[0] if response.embeddings.float_ else []
             else:
-                logger.warning("No embeddings found")
+                log_warning("No embeddings found")
                 return []
         except Exception as e:
-            logger.warning(e)
+            log_warning(e)
             return []
 
     async def async_get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict[str, Any]]]:
@@ -142,3 +249,75 @@ class CohereEmbedder(Embedder):
         if usage:
             return embedding, usage.model_dump()
         return embedding, 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:
+        s, List of usage dictionaries)
+        """
+        all_embeddings = []
+        all_usage = []
+        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]
+
+            try:
+                # Use retry logic for batch processing
+                batch_embeddings, batch_usage = await self._async_batch_with_retry(batch_texts)
+                all_embeddings.extend(batch_embeddings)
+                all_usage.extend(batch_usage)
+
+            except Exception as e:
+                log_warning(f"Async batch embedding failed after retries: {e}")
+
+                # Check if this is a rate limit error and backoff is disabled
+                if self._is_rate_limit_error(e) and not self.exponential_backoff:
+                    log_warning("Rate limit hit and backoff is disabled. Failing immediately.")
+                    raise e
+
+                # Only fall back to individual calls for non-rate-limit errors
+                # For rate limit errors, we should reduce batch size instead
+                if self._is_rate_limit_error(e):
+                    log_warning("Rate limit hit even after retries. Consider reducing batch_size or upgrading API key.")
+                    # Try with smaller batch size
+                    if len(batch_texts) > 1:
+                        smaller_batch_size = max(1, len(batch_texts) // 2)
+                        log_info(f"Retrying with smaller batch size: {smaller_batch_size}")
+                        for j in range(0, len(batch_texts), smaller_batch_size):
+                            small_batch = batch_texts[j : j + smaller_batch_size]
+                            try:
+                                small_embeddings, small_usage = await self._async_batch_with_retry(small_batch)
+                                all_embeddings.extend(small_embeddings)
+                                all_usage.extend(small_usage)
+                            except Exception as e3:
+                                log_error(f"Failed even with reduced batch size: {e3}")
+                                # Fall back to empty results for this batch
+                                all_embeddings.extend([[] for _ in small_batch])
+                                all_usage.extend([None for _ in small_batch])
+                    else:
+                        # Single item already failed, add empty result
+                        log_debug("Single item failed, adding empty result")
+                        all_embeddings.append([])
+                        all_usage.append(None)
+                else:
+                    # For non-rate-limit errors, fall back to individual calls
+                    log_debug("Non-rate-limit error, falling back to individual calls")
+                    for text in batch_texts:
+                        try:
+                            embedding, usage = await self.async_get_embedding_and_usage(text)
+                            all_embeddings.append(embedding)
+                            all_usage.append(usage)
+                        except Exception as e2:
+                            log_warning(f"Error in individual async embedding fallback: {e2}")
+                            all_embeddings.append([])
+                            all_usage.append(None)
+
+        return all_embeddings, all_usage

agno/knowledge/embedder/google.py

@@ -3,7 +3,7 @@ from os import getenv
 from typing import Any, Dict, List, Optional, Tuple
 
 from agno.knowledge.embedder.base import Embedder
-from agno.utils.log import log_error, log_info
+from agno.utils.log import log_error, log_info, log_warning
 
 try:
     from google import genai
@@ -178,3 +178,81 @@ class GeminiEmbedder(Embedder):
         except Exception as e:
             log_error(f"Error extracting embeddings: {e}")
             return [], usage
+
+    async def async_get_embeddings_batch_and_usage(
+        self, texts: List[str]
+    ) -> Tuple[List[List[float]], List[Optional[Dict[str, Any]]]]:
+        """
+        Get embeddings and usage for multiple texts in batches.
+
+        Args:
+            texts: List of text strings to embed
+
+        Returns:
+            Tuple of (List of embedding vectors, List of usage dictionaries)
+        """
+        all_embeddings: List[List[float]] = []
+        all_usage: List[Optional[Dict[str, Any]]] = []
+        log_info(f"Getting embeddings and usage for {len(texts)} texts in batches of {self.batch_size}")
+
+        for i in range(0, len(texts), self.batch_size):
+            batch_texts = texts[i : i + self.batch_size]
+
+            # If a user provides a model id with the `models/` prefix, we need to remove it
+            _id = self.id
+            if _id.startswith("models/"):
+                _id = _id.split("/")[-1]
+
+            _request_params: Dict[str, Any] = {"contents": batch_texts, "model": _id, "config": {}}
+            if self.dimensions:
+                _request_params["config"]["output_dimensionality"] = self.dimensions
+            if self.task_type:
+                _request_params["config"]["task_type"] = self.task_type
+            if self.title:
+                _request_params["config"]["title"] = self.title
+            if not _request_params["config"]:
+                del _request_params["config"]
+
+            if self.request_params:
+                _request_params.update(self.request_params)
+
+            try:
+                response = await self.aclient.aio.models.embed_content(**_request_params)
+
+                # Extract embeddings from batch response
+                if response.embeddings:
+                    batch_embeddings = []
+                    for embedding in response.embeddings:
+                        if embedding.values is not None:
+                            batch_embeddings.append(embedding.values)
+                        else:
+                            batch_embeddings.append([])
+                    all_embeddings.extend(batch_embeddings)
+                else:
+                    # If no embeddings, add empty lists for each text in batch
+                    all_embeddings.extend([[] for _ in batch_texts])
+
+                # Extract usage information
+                usage_dict = None
+                if response.metadata and hasattr(response.metadata, "billable_character_count"):
+                    usage_dict = {"billable_character_count": response.metadata.billable_character_count}
+
+                # Add same usage info for each embedding in the batch
+                all_usage.extend([usage_dict] * len(batch_texts))
+
+            except Exception as e:
+                log_warning(f"Error in async batch embedding: {e}")
+                # Fallback to individual calls for this batch
+                for text in batch_texts:
+                    try:
+                        text_embedding: List[float]
+                        text_usage: Optional[Dict[str, Any]]
+                        text_embedding, text_usage = await self.async_get_embedding_and_usage(text)
+                        all_embeddings.append(text_embedding)
+                        all_usage.append(text_usage)
+                    except Exception as e2:
+                        log_warning(f"Error in individual async embedding fallback: {e2}")
+                        all_embeddings.append([])
+                        all_usage.append(None)
+
+        return all_embeddings, all_usage

agno/knowledge/embedder/huggingface.py

@@ -3,12 +3,12 @@ from os import getenv
 from typing import Any, Dict, List, Optional, Tuple
 
 from agno.knowledge.embedder.base import Embedder
-from agno.utils.log import logger
+from agno.utils.log import log_error, log_warning
 
 try:
     from huggingface_hub import AsyncInferenceClient, InferenceClient
 except ImportError:
-    logger.error("`huggingface-hub` not installed, please run `pip install huggingface-hub`")
+    log_error("`huggingface-hub` not installed, please run `pip install huggingface-hub`")
     raise
 
 
@@ -22,6 +22,11 @@ class HuggingfaceCustomEmbedder(Embedder):
     huggingface_client: Optional[InferenceClient] = None
     async_client: Optional[AsyncInferenceClient] = None
 
+    def __post_init__(self):
+        if self.enable_batch:
+            log_warning("HuggingfaceEmbedder does not support batch embeddings, setting enable_batch to False")
+            self.enable_batch = False
+
     @property
     def client(self) -> InferenceClient:
         if self.huggingface_client:
@@ -61,7 +66,7 @@ class HuggingfaceCustomEmbedder(Embedder):
             else:
                 return list(response)
         except Exception as e:
-            logger.warning(f"Failed to process embeddings: {e}")
+            log_warning(f"Failed to process embeddings: {e}")
             return []
 
     def get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict]]:
@@ -80,7 +85,7 @@ class HuggingfaceCustomEmbedder(Embedder):
             else:
                 return list(response)
         except Exception as e:
-            logger.warning(f"Failed to process embeddings: {e}")
+            log_warning(f"Failed to process embeddings: {e}")
             return []
 
     async def async_get_embedding_and_usage(self, text: str) -> Tuple[List[float], Optional[Dict]]:

agno/knowledge/embedder/jina.py

@@ -117,3 +117,66 @@ class JinaEmbedder(Embedder):
         except Exception as e:
             logger.warning(f"Failed to get embedding and usage: {e}")
             return [], None
+
+    async def _async_batch_response(self, texts: List[str]) -> Dict[str, Any]:
+        """Async batch version of _response using aiohttp."""
+        data = {
+            "model": self.id,
+            "late_chunking": self.late_chunking,
+            "dimensions": self.dimensions,
+            "embedding_type": self.embedding_type,
+            "input": texts,  # Jina API expects a list of texts for batch processing
+        }
+        if self.user is not None:
+            data["user"] = self.user
+        if self.request_params:
+            data.update(self.request_params)
+
+        timeout = aiohttp.ClientTimeout(total=self.timeout) if self.timeout else None
+
+        async with aiohttp.ClientSession(timeout=timeout) as session:
+            async with session.post(self.base_url, headers=self._get_headers(), json=data) as response:
+                response.raise_for_status()
+                return await response.json()
+
+    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.
+
+        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 and usage for {len(texts)} texts in batches of {self.batch_size}")
+
+        for i in range(0, len(texts), self.batch_size):
+            batch_texts = texts[i : i + self.batch_size]
+
+            try:
+                result = await self._async_batch_response(batch_texts)
+                batch_embeddings = [data["embedding"] for data in result["data"]]
+                all_embeddings.extend(batch_embeddings)
+
+                # For each embedding in the batch, add the same usage information
+                usage_dict = result.get("usage")
+                all_usage.extend([usage_dict] * len(batch_embeddings))
+            except Exception as e:
+                logger.warning(f"Error in async batch embedding: {e}")
+                # Fallback to individual calls for this batch
+                for text in batch_texts:
+                    try:
+                        embedding, usage = await self.async_get_embedding_and_usage(text)
+                        all_embeddings.append(embedding)
+                        all_usage.append(usage)
+                    except Exception as e2:
+                        logger.warning(f"Error in individual async embedding fallback: {e2}")
+                        all_embeddings.append([])
+                        all_usage.append(None)
+
+        return all_embeddings, all_usage

agno/knowledge/embedder/mistral.py

@@ -3,13 +3,13 @@ from os import getenv
 from typing import Any, Dict, List, Optional, Tuple
 
 from agno.knowledge.embedder.base import Embedder
-from agno.utils.log import logger
+from agno.utils.log import log_error, log_info, log_warning
 
 try:
     from mistralai import Mistral  # type: ignore
     from mistralai.models.embeddingresponse import EmbeddingResponse  # type: ignore
 except ImportError:
-    logger.error("`mistralai` not installed")
+    log_error("`mistralai` not installed")
     raise
 
 
@@ -50,7 +50,7 @@ class MistralEmbedder(Embedder):
 
     def _response(self, text: str) -> EmbeddingResponse:
         _request_params: Dict[str, Any] = {
-            "inputs": text,
+            "inputs": [text],  # Mistral API expects a list
             "model": self.id,
         }
         if self.request_params:
@@ -67,7 +67,7 @@ class MistralEmbedder(Embedder):
                 return response.data[0].embedding
             return []
         except Exception as e:
-            logger.warning(f"Error getting embedding: {e}")
+            log_warning(f"Error getting embedding: {e}")
             return []
 
     def get_embedding_and_usage(self, text: str) -> Tuple[List[float], Dict[str, Any]]:
@@ -79,7 +79,7 @@ class MistralEmbedder(Embedder):
             usage: Dict[str, Any] = response.usage.model_dump() if response.usage else {}
             return embedding, usage
         except Exception as e:
-            logger.warning(f"Error getting embedding and usage: {e}")
+            log_warning(f"Error getting embedding and usage: {e}")
             return [], {}
 
     async def async_get_embedding(self, text: str) -> List[float]:
@@ -88,7 +88,7 @@ class MistralEmbedder(Embedder):
             # Check if the client has an async version of embeddings.create
             if hasattr(self.client.embeddings, "create_async"):
                 response: EmbeddingResponse = await self.client.embeddings.create_async(
-                    inputs=text, model=self.id, **self.request_params if self.request_params else {}
+                    inputs=[text], model=self.id, **self.request_params if self.request_params else {}
                 )
             else:
                 # Fallback to running sync method in thread executor
@@ -98,7 +98,7 @@ class MistralEmbedder(Embedder):
                 response: EmbeddingResponse = await loop.run_in_executor(  # type: ignore
                     None,
                     lambda: self.client.embeddings.create(
-                        inputs=text, model=self.id, **self.request_params if self.request_params else {}
+                        inputs=[text], model=self.id, **self.request_params if self.request_params else {}
                     ),
                 )
 
@@ -106,7 +106,7 @@ class MistralEmbedder(Embedder):
                 return response.data[0].embedding
             return []
         except Exception as e:
-            logger.warning(f"Error getting embedding: {e}")
+            log_warning(f"Error getting embedding: {e}")
             return []
 
     async def async_get_embedding_and_usage(self, text: str) -> Tuple[List[float], Dict[str, Any]]:
@@ -115,7 +115,7 @@ class MistralEmbedder(Embedder):
             # Check if the client has an async version of embeddings.create
             if hasattr(self.client.embeddings, "create_async"):
                 response: EmbeddingResponse = await self.client.embeddings.create_async(
-                    inputs=text, model=self.id, **self.request_params if self.request_params else {}
+                    inputs=[text], model=self.id, **self.request_params if self.request_params else {}
                 )
             else:
                 # Fallback to running sync method in thread executor
@@ -125,7 +125,7 @@ class MistralEmbedder(Embedder):
                 response: EmbeddingResponse = await loop.run_in_executor(  # type: ignore
                     None,
                     lambda: self.client.embeddings.create(
-                        inputs=text, model=self.id, **self.request_params if self.request_params else {}
+                        inputs=[text], model=self.id, **self.request_params if self.request_params else {}
                     ),
                 )
 
@@ -135,5 +135,72 @@ class MistralEmbedder(Embedder):
             usage: Dict[str, Any] = response.usage.model_dump() if response.usage else {}
             return embedding, usage
         except Exception as e:
-            logger.warning(f"Error getting embedding and usage: {e}")
+            log_warning(f"Error getting embedding and usage: {e}")
             return [], {}
+
+    async def async_get_embeddings_batch_and_usage(
+        self, texts: List[str]
+    ) -> Tuple[List[List[float]], List[Optional[Dict[str, Any]]]]:
+        """
+        Get embeddings and usage for multiple texts in batches.
+
+        Args:
+            texts: List of text strings to embed
+
+        Returns:
+            Tuple of (List of embedding vectors, List of usage dictionaries)
+        """
+        all_embeddings = []
+        all_usage = []
+        log_info(f"Getting embeddings and usage for {len(texts)} texts in batches of {self.batch_size}")
+
+        for i in range(0, len(texts), self.batch_size):
+            batch_texts = texts[i : i + self.batch_size]
+
+            _request_params: Dict[str, Any] = {
+                "inputs": batch_texts,  # Mistral API expects a list for batch processing
+                "model": self.id,
+            }
+            if self.request_params:
+                _request_params.update(self.request_params)
+
+            try:
+                # Check if the client has an async version of embeddings.create
+                if hasattr(self.client.embeddings, "create_async"):
+                    response: EmbeddingResponse = await self.client.embeddings.create_async(**_request_params)
+                else:
+                    # Fallback to running sync method in thread executor
+                    import asyncio
+
+                    loop = asyncio.get_running_loop()
+                    response: EmbeddingResponse = await loop.run_in_executor(  # type: ignore
+                        None, lambda: self.client.embeddings.create(**_request_params)
+                    )
+
+                # Extract embeddings from batch response
+                if response.data:
+                    batch_embeddings = [data.embedding for data in response.data if data.embedding]
+                    all_embeddings.extend(batch_embeddings)
+                else:
+                    # If no embeddings, add empty lists for each text in batch
+                    all_embeddings.extend([[] for _ in batch_texts])
+
+                # Extract usage information
+                usage_dict = response.usage.model_dump() if response.usage else None
+                # Add same usage info for each embedding in the batch
+                all_usage.extend([usage_dict] * len(batch_texts))
+
+            except Exception as e:
+                log_warning(f"Error in async batch embedding: {e}")
+                # Fallback to individual calls for this batch
+                for text in batch_texts:
+                    try:
+                        embedding, usage = await self.async_get_embedding_and_usage(text)
+                        all_embeddings.append(embedding)
+                        all_usage.append(usage)
+                    except Exception as e2:
+                        log_warning(f"Error in individual async embedding fallback: {e2}")
+                        all_embeddings.append([])
+                        all_usage.append(None)
+
+        return all_embeddings, all_usage

agno/knowledge/embedder/ollama.py

@@ -45,6 +45,11 @@ class OllamaEmbedder(Embedder):
     ollama_client: Optional[OllamaClient] = None
     async_client: Optional[AsyncOllamaClient] = None
 
+    def __post_init__(self):
+        if self.enable_batch:
+            logger.warning("OllamaEmbedder does not support batch embeddings, setting enable_batch to False")
+            self.enable_batch = False
+
     @property
     def client(self) -> OllamaClient:
         if self.ollama_client: