haystack-experimental 0.13.0__py3-none-any.whl → 0.14.1__py3-none-any.whl

haystack_experimental/components/summarizers/llm_summarizer.py

@@ -0,0 +1,317 @@
+# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Any, Optional
+
+from haystack import Document, component, default_from_dict, default_to_dict, logging
+from haystack.components.generators.chat.types import ChatGenerator
+from haystack.components.preprocessors import RecursiveDocumentSplitter
+from haystack.core.serialization import component_to_dict
+from haystack.dataclasses import ChatMessage
+from haystack.utils import deserialize_chatgenerator_inplace
+from tqdm import tqdm
+
+logger = logging.getLogger(__name__)
+
+
+@component
+class LLMSummarizer:
+    """
+    Summarizes text using a language model.
+
+    It's inspired by code from the OpenAI blog post: https://cookbook.openai.com/examples/summarizing_long_documents
+
+    Example
+    ```python
+    from haystack_experimental.components.summarizers.summarizer import Summarizer
+    from haystack.components.generators.chat import OpenAIChatGenerator
+    from haystack import Document
+
+    text = ("Machine learning is a subset of artificial intelligence that provides systems "
+            "the ability to automatically learn and improve from experience without being "
+            "explicitly programmed. The process of learning begins with observations or data. "
+            "Supervised learning algorithms build a mathematical model of sample data, known as "
+            "training data, in order to make predictions or decisions. Unsupervised learning "
+            "algorithms take a set of data that contains only inputs and find structure in the data. "
+            "Reinforcement learning is an area of machine learning where an agent learns to behave "
+            "in an environment by performing actions and seeing the results. Deep learning uses "
+            "artificial neural networks to model complex patterns in data. Neural networks consist "
+            "of layers of connected nodes, each performing a simple computation.")
+
+    doc = Document(content=text)
+    chat_generator = OpenAIChatGenerator(model="gpt-4")
+    summarizer = Summarizer(chat_generator=chat_generator)
+    summarizer.run(documents=[doc])
+    ```
+    """
+
+    def __init__(  # pylint: disable=too-many-positional-arguments
+        self,
+        chat_generator: ChatGenerator,
+        system_prompt: Optional[str] = "Rewrite this text in summarized form.",
+        summary_detail: float = 0,
+        minimum_chunk_size: Optional[int] = 500,
+        chunk_delimiter: str = ".",
+        summarize_recursively: bool = False,
+        split_overlap: int = 0,
+    ):
+        """
+        Initialize the Summarizer component.
+
+        :param chat_generator: A ChatGenerator instance to use for summarization.
+        :param system_prompt: The prompt to instruct the LLM to summarise text, if not given defaults to:
+            "Rewrite this text in summarized form."
+        :param summary_detail: The level of detail for the summary (0-1), defaults to 0.
+            This parameter controls the trade-off between conciseness and completeness by adjusting how many
+            chunks the text is divided into. At detail=0, the text is processed as a single chunk (or very few
+            chunks), producing the most concise summary. At detail=1, the text is split into the maximum number
+            of chunks allowed by minimum_chunk_size, enabling more granular analysis and detailed summaries.
+            The formula uses linear interpolation: num_chunks = 1 + detail * (max_chunks - 1), where max_chunks
+            is determined by dividing the document length by minimum_chunk_size.
+        :param minimum_chunk_size: The minimum token count per chunk, defaults to 500
+        :param chunk_delimiter: The character used to determine separator priority.
+            "." uses sentence-based splitting, "\n" uses paragraph-based splitting, defaults to "."
+        :param summarize_recursively: Whether to use previous summaries as context, defaults to False.
+        :param split_overlap: Number of tokens to overlap between consecutive chunks, defaults to 0.
+        """
+        self._chat_generator = chat_generator
+        self.summary_detail = summary_detail
+        self.minimum_chunk_size = minimum_chunk_size
+        self.chunk_delimiter = chunk_delimiter
+        self.system_prompt = system_prompt
+        self.summarize_recursively = summarize_recursively
+        self.split_overlap = split_overlap
+
+        # Map chunk_delimiter to an appropriate separator strategy
+        separators = LLMSummarizer._get_separators_from_delimiter(chunk_delimiter)
+
+        # Initialize RecursiveDocumentSplitter
+        # Note: split_length will be updated dynamically based on detail parameter
+        self._document_splitter = RecursiveDocumentSplitter(
+            split_length=minimum_chunk_size if minimum_chunk_size else 500,
+            split_overlap=split_overlap,
+            split_unit="token",
+            separators=separators,
+        )
+
+    @staticmethod
+    def _get_separators_from_delimiter(delimiter: str) -> list[str]:
+        """
+        Map the delimiter to an appropriate list of separators for RecursiveDocumentSplitter.
+
+        :param delimiter: The delimiter character
+        :returns: List of separators in order of preference
+        """
+        if delimiter == ".":
+            # Sentence-focused splitting
+            return ["\n\n", "sentence", "\n", " "]
+        elif delimiter == "\n":
+            # Paragraph-focused splitting
+            return ["\n\n", "\n", "sentence", " "]
+        else:
+            # Custom delimiter - prioritize it
+            return ["\n\n", delimiter, "\n", " "]
+
+    def warm_up(self):
+        """
+        Warm up the chat generator and document splitter components.
+        """
+        # Warm up chat generator
+        if hasattr(self._chat_generator, "warm_up"):
+            self._chat_generator.warm_up()
+
+        # Warm up document splitter (needed for sentence splitting and tokenization)
+        if hasattr(self._document_splitter, "warm_up"):
+            self._document_splitter.warm_up()
+
+    def to_dict(self) -> dict[str, Any]:
+        """
+        Serializes the component to a dictionary.
+
+        :returns:
+            Dictionary with serialized data.
+        """
+        return default_to_dict(
+            self,
+            chat_generator=component_to_dict(obj=self._chat_generator, name="chat_generator"),
+            system_prompt=self.system_prompt,
+            summary_detail=self.summary_detail,
+            minimum_chunk_size=self.minimum_chunk_size,
+            chunk_delimiter=self.chunk_delimiter,
+            summarize_recursively=self.summarize_recursively,
+            split_overlap=self.split_overlap,
+        )
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> "LLMSummarizer":
+        """
+        Deserializes the component from a dictionary.
+
+        :param data: Dictionary with serialized data.
+        :returns:
+            An instance of the component.
+        """
+        init_params = data.get("init_parameters", {})
+
+        # Deserialize chat_generator
+        deserialize_chatgenerator_inplace(init_params, key="chat_generator")
+
+        return default_from_dict(cls, data)
+
+    def num_tokens(self, text: str) -> int:
+        """
+        Estimates the token count for a given text.
+
+        Uses the RecursiveDocumentSplitter's tokenization logic for consistency.
+
+        :param text: The text to tokenize
+        :returns:
+            The estimated token count
+        """
+        # Use the document splitter's tokenization method for consistency
+        return self._document_splitter._chunk_length(text)
+
+    def _prepare_text_chunks(self, text, detail, minimum_chunk_size, chunk_delimiter):
+        """
+        Prepares text chunks based on detail level using RecursiveDocumentSplitter.
+
+        The detail parameter (0-1) controls the granularity through linear interpolation:
+        - detail=0: Creates fewer, larger chunks (most concise summary)
+        - detail=1: Creates more, smaller chunks (most detailed summary)
+
+        The formula calculates: num_chunks = 1 + detail * (max_chunks - 1), where max_chunks is the
+        document_length divided by minimum_chunk_size. This interpolates between processing the entire
+        text as one chunk (detail=0) and splitting it into the maximum number of chunks that respect
+        the minimum_chunk_size constraint (detail=1). Higher detail values allow the LLM to analyze
+        smaller portions of text more carefully, preserving more information at the cost of longer
+        processing time and potentially longer summaries.
+
+        :param text: The text to chunk
+        :param detail: Detail level (0-1)
+        :param minimum_chunk_size: Minimum token count per chunk
+        :param chunk_delimiter: Delimiter for separator selection
+        :returns: List of text chunks
+        """
+        document_length = self.num_tokens(text)
+        max_chunks = max(1, document_length // minimum_chunk_size)
+        min_chunks = 1
+
+        num_chunks = int(min_chunks + detail * (max_chunks - min_chunks))
+        num_chunks = max(1, num_chunks)  # Ensure at least 1 chunk
+
+        chunk_size = max(minimum_chunk_size, document_length // num_chunks)
+        self._document_splitter.split_length = chunk_size
+        if chunk_delimiter != self.chunk_delimiter:
+            self._document_splitter.separators = self._get_separators_from_delimiter(chunk_delimiter)
+
+        temp_doc = Document(content=text)
+        result = self._document_splitter.run(documents=[temp_doc])
+        text_chunks = [doc.content for doc in result["documents"]]
+
+        return text_chunks
+
+    def _process_chunks(self, text_chunks, summarize_recursively):
+        """
+        Processes each chunk individually, asking the LLM to summarize it, and accumulates all the summaries.
+
+        The parameter `summarize_recursively` allows to use previous summaries as context for the next chunk.
+        """
+        accumulated_summaries: list[str] = []
+
+        for chunk in tqdm(text_chunks):
+            if summarize_recursively and accumulated_summaries:
+                accumulated_summaries_string = "\n\n".join(accumulated_summaries)
+                user_message_content = (
+                    f"Previous summaries:\n\n{accumulated_summaries_string}\n\nText to summarize next:\n\n{chunk}"
+                )
+            else:
+                user_message_content = chunk
+
+            # prepare the message and make the LLM call
+            # self.system_prompt is not None where due to the default value in the constructor
+            messages = [ChatMessage.from_system(self.system_prompt), ChatMessage.from_user(user_message_content)]  # type: ignore
+            result = self._chat_generator.run(messages=messages)
+            accumulated_summaries.append(result["replies"][0].text)
+
+        return accumulated_summaries
+
+    def summarize(
+        self,
+        text: str,
+        detail: float,
+        minimum_chunk_size: int,
+        summarize_recursively: bool = False,
+    ) -> str:
+        """
+        Summarizes text by splitting it into optimally-sized chunks and processing each with an LLM.
+
+        :param text: Text to summarize
+        :param detail: Detail level (0-1) where 0 is most concise and 1 is most detailed
+        :param minimum_chunk_size: Minimum token count per chunk
+        :param summarize_recursively: Whether to use previous summaries as context
+
+        :returns:
+            The textual content summarized by the LLM.
+
+        :raises ValueError: If detail is not between 0 and 1
+        """
+
+        if not 0 <= detail <= 1:
+            raise ValueError("Detail must be between 0 and 1")
+
+        # calculate "optimal" chunking parameters
+        text_chunks = self._prepare_text_chunks(text, detail, minimum_chunk_size, self.chunk_delimiter)
+
+        # process chunks and accumulate summaries
+        accumulated_summaries = self._process_chunks(text_chunks, summarize_recursively)
+
+        # combine all summaries
+        return "\n\n".join(accumulated_summaries)
+
+    @component.output_types(summary=list[Document])
+    def run(
+        self,
+        *,
+        documents: list[Document],
+        detail: Optional[float] = None,
+        minimum_chunk_size: Optional[int] = None,
+        summarize_recursively: Optional[bool] = None,
+        system_prompt: Optional[str] = None,
+    ) -> dict[str, list[Document]]:
+        """
+        Run the summarizer on a list of documents.
+
+        :param documents: List of documents to summarize
+        :param detail: The level of detail for the summary (0-1), defaults to 0 overwriting the component's default.
+        :param minimum_chunk_size: The minimum token count per chunk, defaults to 500 overwriting the
+                                   component's default.
+        :param system_prompt: If given it will overwrite prompt given at init time or the default one.
+        :param summarize_recursively: Whether to use previous summaries as context, defaults to False overwriting the
+                                      component's default.
+
+        :raises RuntimeError: If the component wasn't warmed up.
+        """
+
+        if not self._document_splitter._is_warmed_up:
+            raise RuntimeError("The Summarizer component wasn't warmed up. Call 'warm_up()' before calling 'run()'.")
+
+        # let's allow to change some of the parameters at run time
+        detail = self.summary_detail if detail is None else detail
+        minimum_chunk_size = self.minimum_chunk_size if minimum_chunk_size is None else minimum_chunk_size
+        summarize_recursively = self.summarize_recursively if summarize_recursively is None else summarize_recursively
+        self.system_prompt = system_prompt if system_prompt else self.system_prompt
+
+        for doc in documents:
+            if doc.content is None or doc.content == "":
+                logger.warning("Document ID {doc_id} has an empty content. Skipping this document.", doc_id=doc.id)
+                continue
+            summary = self.summarize(
+                doc.content,
+                detail=detail,
+                minimum_chunk_size=minimum_chunk_size,  # type: ignore # already checked, cannot be None here
+                summarize_recursively=summarize_recursively,
+            )
+            doc.meta["summary"] = summary
+
+        return {"documents": documents}

haystack_experimental/core/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
+#
+# SPDX-License-Identifier: Apache-2.0

haystack_experimental/core/pipeline/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
+#
+# SPDX-License-Identifier: Apache-2.0

haystack_experimental/core/pipeline/breakpoint.py

@@ -0,0 +1,119 @@
+# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from copy import deepcopy
+from dataclasses import replace
+from datetime import datetime
+from typing import TYPE_CHECKING, Any, Optional
+
+from haystack import logging
+from haystack.dataclasses.breakpoints import AgentBreakpoint, PipelineSnapshot, PipelineState, ToolBreakpoint
+from haystack.utils.base_serialization import _serialize_value_with_schema
+from haystack.utils.misc import _get_output_dir
+
+from haystack_experimental.dataclasses.breakpoints import AgentSnapshot
+
+if TYPE_CHECKING:
+    from haystack_experimental.components.agents.agent import _ExecutionContext
+    from haystack_experimental.components.agents.human_in_the_loop import ToolExecutionDecision
+
+
+logger = logging.getLogger(__name__)
+
+
+def _create_agent_snapshot(
+    *,
+    component_visits: dict[str, int],
+    agent_breakpoint: AgentBreakpoint,
+    component_inputs: dict[str, Any],
+    tool_execution_decisions: Optional[list["ToolExecutionDecision"]] = None,
+) -> AgentSnapshot:
+    """
+    Create a snapshot of the agent's state.
+
+    NOTE: Only difference to Haystack's native implementation is the addition of tool_execution_decisions to the
+    AgentSnapshot.
+
+    :param component_visits: The visit counts for the agent's components.
+    :param agent_breakpoint: AgentBreakpoint object containing breakpoints
+    :param component_inputs: The inputs to the agent's components.
+    :param tool_execution_decisions: Optional list of ToolExecutionDecision objects representing decisions made
+        regarding tool executions.
+    :return: An AgentSnapshot containing the agent's state and component visits.
+    """
+    return AgentSnapshot(
+        component_inputs={
+            "chat_generator": _serialize_value_with_schema(deepcopy(component_inputs["chat_generator"])),
+            "tool_invoker": _serialize_value_with_schema(deepcopy(component_inputs["tool_invoker"])),
+        },
+        component_visits=component_visits,
+        break_point=agent_breakpoint,
+        timestamp=datetime.now(),
+        tool_execution_decisions=tool_execution_decisions,
+    )
+
+
+def _create_pipeline_snapshot_from_tool_invoker(
+    *,
+    execution_context: "_ExecutionContext",
+    tool_name: Optional[str] = None,
+    agent_name: Optional[str] = None,
+    break_point: Optional[AgentBreakpoint] = None,
+    parent_snapshot: Optional[PipelineSnapshot] = None,
+) -> PipelineSnapshot:
+    """
+    Create a pipeline snapshot when a tool invoker breakpoint is raised or an exception during execution occurs.
+
+    :param execution_context: The current execution context of the agent.
+    :param tool_name: The name of the tool that triggered the breakpoint, if available.
+    :param agent_name: The name of the agent component if present in a pipeline.
+    :param break_point: An optional AgentBreakpoint object. If provided, it will be used instead of creating a new one.
+        A scenario where a new breakpoint is created is when an exception occurs during tool execution and we want to
+        capture the state at that point.
+    :param parent_snapshot: An optional parent PipelineSnapshot to build upon.
+    :returns:
+        A PipelineSnapshot containing the state of the pipeline and agent at the point of the breakpoint or exception.
+    """
+    if break_point is None:
+        agent_breakpoint = AgentBreakpoint(
+            agent_name=agent_name or "agent",
+            break_point=ToolBreakpoint(
+                component_name="tool_invoker",
+                visit_count=execution_context.component_visits["tool_invoker"],
+                tool_name=tool_name,
+                snapshot_file_path=_get_output_dir("pipeline_snapshot"),
+            ),
+        )
+    else:
+        agent_breakpoint = break_point
+
+    messages = execution_context.state.data["messages"]
+    agent_snapshot = _create_agent_snapshot(
+        component_visits=execution_context.component_visits,
+        agent_breakpoint=agent_breakpoint,
+        component_inputs={
+            "chat_generator": {"messages": messages[:-1], **execution_context.chat_generator_inputs},
+            "tool_invoker": {
+                "messages": messages[-1:],  # tool invoker consumes last msg from the chat_generator, contains tool call
+                "state": execution_context.state,
+                **execution_context.tool_invoker_inputs,
+            },
+        },
+        tool_execution_decisions=execution_context.tool_execution_decisions,
+    )
+    if parent_snapshot is None:
+        # Create an empty pipeline snapshot if no parent snapshot is provided
+        final_snapshot = PipelineSnapshot(
+            pipeline_state=PipelineState(inputs={}, component_visits={}, pipeline_outputs={}),
+            timestamp=agent_snapshot.timestamp,
+            break_point=agent_snapshot.break_point,
+            agent_snapshot=agent_snapshot,
+            original_input_data={},
+            ordered_component_names=[],
+            include_outputs_from=set(),
+        )
+    else:
+        final_snapshot = replace(parent_snapshot, agent_snapshot=agent_snapshot)
+
+    return final_snapshot

haystack_experimental/dataclasses/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
+#
+# SPDX-License-Identifier: Apache-2.0

haystack_experimental/dataclasses/breakpoints.py

@@ -0,0 +1,53 @@
+# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from dataclasses import dataclass
+from datetime import datetime
+from typing import Any, Optional
+
+from haystack.dataclasses.breakpoints import AgentBreakpoint
+from haystack.dataclasses.breakpoints import AgentSnapshot as HaystackAgentSnapshot
+
+from haystack_experimental.components.agents.human_in_the_loop.dataclasses import ToolExecutionDecision
+
+
+@dataclass
+class AgentSnapshot(HaystackAgentSnapshot):
+    tool_execution_decisions: Optional[list[ToolExecutionDecision]] = None
+
+    def to_dict(self) -> dict[str, Any]:
+        """
+        Convert the AgentSnapshot to a dictionary representation.
+
+        :return: A dictionary containing the agent state, timestamp, and breakpoint.
+        """
+        return {
+            "component_inputs": self.component_inputs,
+            "component_visits": self.component_visits,
+            "break_point": self.break_point.to_dict(),
+            "timestamp": self.timestamp.isoformat() if self.timestamp else None,
+            "tool_execution_decisions": [ted.to_dict() for ted in self.tool_execution_decisions]
+            if self.tool_execution_decisions
+            else None,
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "AgentSnapshot":
+        """
+        Populate the AgentSnapshot from a dictionary representation.
+
+        :param data: A dictionary containing the agent state, timestamp, and breakpoint.
+        :return: An instance of AgentSnapshot.
+        """
+        return cls(
+            component_inputs=data["component_inputs"],
+            component_visits=data["component_visits"],
+            break_point=AgentBreakpoint.from_dict(data["break_point"]),
+            timestamp=datetime.fromisoformat(data["timestamp"]) if data.get("timestamp") else None,
+            tool_execution_decisions=[
+                ToolExecutionDecision.from_dict(ted) for ted in data.get("tool_execution_decisions", [])
+            ]
+            if data.get("tool_execution_decisions")
+            else None,
+        )

{haystack_experimental-0.13.0.dist-info → haystack_experimental-0.14.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: haystack-experimental
-Version: 0.13.0
+Version: 0.14.1
 Summary: Experimental components and features for the Haystack LLM framework.
 Project-URL: CI: GitHub, https://github.com/deepset-ai/haystack-experimental/actions
 Project-URL: GitHub: issues, https://github.com/deepset-ai/haystack-experimental/issues
@@ -25,6 +25,7 @@ Classifier: Programming Language :: Python :: 3.13
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
 Requires-Dist: haystack-ai
+Requires-Dist: rich
 Description-Content-Type: text/markdown
 
 [![PyPI - Version](https://img.shields.io/pypi/v/haystack-experimental.svg)](https://pypi.org/project/haystack-experimental)
@@ -70,16 +71,19 @@ that includes it. Once it reaches the end of its lifespan, the experiment will b
 
 ### Active experiments
 
-| Name                                  | Type                           | Expected End Date | Dependencies | Cookbook                                                                                                                                                                                                                                                  | Discussion    |
-|---------------------------------------|--------------------------------|-------------------|--------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------|
-| [`InMemoryChatMessageStore`][1]       | Memory Store                   | December 2024     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>      | [Discuss][4]  |
-| [`ChatMessageRetriever`][2]           | Memory Component               | December 2024     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>      | [Discuss][4]  |
-| [`ChatMessageWriter`][3]              | Memory Component               | December 2024     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>      | [Discuss][4]  |
-| [`QueryExpander`][5]                  | Query Expansion Component      | October 2025      | None         | None | [Discuss][6]  |
-| [`EmbeddingBasedDocumentSplitter`][8] | EmbeddingBasedDocumentSplitter | August 2025       | None         | None | [Discuss][7]  |
-| [`MultiQueryEmbeddingRetriever`][9]   | MultiQueryEmbeddingRetriever   | November 2025     | None         | None | [Discuss][11] |
-| [`MultiQueryTextRetriever`][10]       | MultiQueryTextRetriever        | November 2025     | None         | None | [Discuss][12] |
-| [`OpenAIChatGenerator`][9]            | Chat Generator Component       | November 2025     | None         | None | [Discuss][10] |
+| Name                                                                                                                                                                                    | Type                           | Expected End Date | Dependencies | Cookbook                                                                                                                                                                                                                                             | Discussion    |
+|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------|-------------------|--------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------|
+| [`InMemoryChatMessageStore`][1]                                                                                                                                                         | Memory Store                   | December 2024     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> | [Discuss][4]  |
+| [`ChatMessageRetriever`][2]                                                                                                                                                             | Memory Component               | December 2024     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> | [Discuss][4]  |
+| [`ChatMessageWriter`][3]                                                                                                                                                                | Memory Component               | December 2024     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/conversational_rag_using_memory.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> | [Discuss][4]  |
+| [`QueryExpander`][5]                                                                                                                                                                    | Query Expansion Component      | October 2025      | None         | None                                                                                                                                                                                                                                                 | [Discuss][6]  |
+| [`EmbeddingBasedDocumentSplitter`][8]                                                                                                                                                   | EmbeddingBasedDocumentSplitter | August 2025       | None         | None                                                                                                                                                                                                                                                 | [Discuss][7]  |
+| [`MultiQueryEmbeddingRetriever`][13]                                                                                                                                                    | MultiQueryEmbeddingRetriever   | November 2025     | None         | None                                                                                                                                                                                                                                                 | [Discuss][11] |
+| [`MultiQueryTextRetriever`][14]                                                                                                                                                         | MultiQueryTextRetriever        | November 2025     | None         | None                                                                                                                                                                                                                                                 | [Discuss][12] |
+| [`OpenAIChatGenerator`][9]                                                                                                                                                              | Chat Generator Component       | November 2025     | None         | <a href="https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/hallucination_score_calculator.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>  | [Discuss][10] |
+| [`MarkdownHeaderLevelInferrer`][15]                                                                                                                                                     | Preprocessor                   | January 2025      | None         | None                                                                                                                                                                                                                                                 | [Discuss][16] |
+| [`Agent`][17]; [Confirmation Policies][18]; [ConfirmationUIs][19]; [ConfirmationStrategies][20]; [`ConfirmationUIResult` and `ToolExecutionDecision`][21] [HITLBreakpointException][22] | Human in the Loop              | December 2025     | rich         | None                                                                                                                                                                                                                                                 | [Discuss][23] |
+| [`LLMSummarizer`][24]                                                                                                                                                                   | Document Summarizer            | January 2025      | None         | None                                                                                                                                                                                                                                                 | [Discuss][25] | 
 
 [1]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/chat_message_stores/in_memory.py
 [2]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/retrievers/chat_message_retriever.py
@@ -90,11 +94,22 @@ that includes it. Once it reaches the end of its lifespan, the experiment will b
 [7]: https://github.com/deepset-ai/haystack-experimental/discussions/356
 [8]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/preprocessors/embedding_based_document_splitter.py
 [9]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/generators/chat/openai.py
-[10]: https://github.com/deepset-ai/haystack-experimental/discussions/XXX
-[11]: https://github.com/deepset-ai/haystack-experimental/discussions/<>
-[12]: https://github.com/deepset-ai/haystack-experimental/discussions/<>
+[10]: https://github.com/deepset-ai/haystack-experimental/discussions/361
+[11]: https://github.com/deepset-ai/haystack-experimental/discussions/363
+[12]: https://github.com/deepset-ai/haystack-experimental/discussions/364
 [13]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/retrievers/multi_query_embedding_retriever.py
 [14]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/retrievers/multi_query_text_retriever.py
+[15]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/preprocessors/md_header_level_inferrer.py      
+[16]: https://github.com/deepset-ai/haystack-experimental/discussions/376
+[17]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/agents/agent.py
+[18]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/agents/human_in_the_loop/policies.py
+[19]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/agents/human_in_the_loop/user_interfaces.py
+[20]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/agents/human_in_the_loop/strategies.py
+[21]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/agents/human_in_the_loop/dataclasses.py
+[22]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/agents/human_in_the_loop/errors.py
+[23]: https://github.com/deepset-ai/haystack-experimental/discussions/381
+[24]: https://github.com/deepset-ai/haystack-experimental/blob/main/haystack_experimental/components/sumarizers/llm_summarizer.py
+[25]: https://github.com/deepset-ai/haystack-experimental/discussions/382
 
 ### Adopted experiments
 | Name                                                                                   | Type                                     | Final release |

{haystack_experimental-0.13.0.dist-info → haystack_experimental-0.14.1.dist-info}/RECORD

@@ -3,14 +3,25 @@ haystack_experimental/chat_message_stores/__init__.py,sha256=sQhjMoaVfC0XsL257eg
 haystack_experimental/chat_message_stores/in_memory.py,sha256=nc_B_70KOvtgsul4QJb7HihHZrO220HRFsm-aXjYFJk,2305
 haystack_experimental/chat_message_stores/types.py,sha256=QzjA5-A08PlMAy7MMLNNUpob8S60Ypec74gSbz_l49E,2101
 haystack_experimental/components/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
+haystack_experimental/components/agents/__init__.py,sha256=Sxu9LxPpQ5cljgoTgUeNC0GY8CwUdiSy1JWkd_-RRJ4,414
+haystack_experimental/components/agents/agent.py,sha256=jJ5baKFgT4PNuf2vu2dlm5AyJ9O7stlK94_1zk2u2OA,31814
+haystack_experimental/components/agents/human_in_the_loop/__init__.py,sha256=xLr1G9pNWMmCpKN9mbv6yqeFfwMcbZyaVfCkzlwMxhY,1674
+haystack_experimental/components/agents/human_in_the_loop/breakpoint.py,sha256=GhNdGdFNDnwSiTukD4WVp6-1YgGjq5oqCEcGMC2dcog,2902
+haystack_experimental/components/agents/human_in_the_loop/dataclasses.py,sha256=OakB0PXBSG0LbQixcuo-d7IC-A3_k6qi80pB8hwY23o,2563
+haystack_experimental/components/agents/human_in_the_loop/errors.py,sha256=HAjD_MCOTBirqnJdxpc2MhqIm-XnU3Soev29wRBWoMw,1066
+haystack_experimental/components/agents/human_in_the_loop/policies.py,sha256=nzblePptT4Fg2GFHa4_SDIK_d7hZ_70qPhkteZBRXWk,3172
+haystack_experimental/components/agents/human_in_the_loop/strategies.py,sha256=KXfMLky27EuxOUhqbHO7oQ3KkL_3lzkwvk1Gk1EMXYY,19643
+haystack_experimental/components/agents/human_in_the_loop/types.py,sha256=aY93Wsd-5BgOiJaaSEGB_bGC-BTx_V_UT1faXtbNNdo,3072
+haystack_experimental/components/agents/human_in_the_loop/user_interfaces.py,sha256=HlJ3-CYNrQGsHOtpvrQE4ayQls8Q3EkLFUkOoRnLVC4,8707
 haystack_experimental/components/embedders/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
 haystack_experimental/components/embedders/types/__init__.py,sha256=HGR8aavwIEx7v-8nm5JxFIw47EWn7vAUmywhakTNDCo,182
 haystack_experimental/components/embedders/types/protocol.py,sha256=EEVtggoYWZL6zF-vbasJollCxLbheMYIISAh7hJ8LkA,1038
 haystack_experimental/components/generators/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
 haystack_experimental/components/generators/chat/__init__.py,sha256=LEKI1mMtltVbSiU40QgBfnWC-z3_660TWuV-cVHhdTw,465
-haystack_experimental/components/generators/chat/openai.py,sha256=QFb-l_VbTkqi46rqE6rkmMoRkT3fX8kEsStOEGMEBRw,10040
+haystack_experimental/components/generators/chat/openai.py,sha256=gX6UI4yfY0pzKhWErquvPF_gV-3Ut0y6wSJytAD07Jk,9855
 haystack_experimental/components/preprocessors/__init__.py,sha256=x3fM1lpGzYjWB3hpdbDWxXr_rYASb2e9yX0PgYG84rA,518
-haystack_experimental/components/preprocessors/embedding_based_document_splitter.py,sha256=moA7brOJHCsZ2FX7RHJp7Miy6gNQWWyMnsC6KE8MYOQ,16540
+haystack_experimental/components/preprocessors/embedding_based_document_splitter.py,sha256=VyQ--gaMsWid-IRBVXi5YPJpwbFlaK-2mRFvRF8MSBQ,17616
+haystack_experimental/components/preprocessors/md_header_level_inferrer.py,sha256=1Tn-H4Gvg2yYSUc54cPWKTCK78KXet5u32_1S8PM3NU,5643
 haystack_experimental/components/query/__init__.py,sha256=quaqe16cbtgIdJx7d56CMdk1zZQ6f_3_TICsU0HF_U8,446
 haystack_experimental/components/query/query_expander.py,sha256=zc9i2zP3ciOWWr029wO_lw3Tl8W3_kQcYcoHDs5Nj8c,12293
 haystack_experimental/components/retrievers/__init__.py,sha256=CqPvqyvGp5L3Y1gTVQC8DD_xHzbIfTzGlj3oCsZM3J8,528
@@ -19,8 +30,15 @@ haystack_experimental/components/retrievers/multi_query_embedding_retriever.py,s
 haystack_experimental/components/retrievers/multi_query_text_retriever.py,sha256=5awo60jy0attihYsNCG4LlJGROSjP2mFX2wAQiXYYOc,6557
 haystack_experimental/components/retrievers/types/__init__.py,sha256=iOngs3gs5enY8y6AWGeyQANTB_9qpXQ0QHSFFDDeEGc,218
 haystack_experimental/components/retrievers/types/protocol.py,sha256=oUdX_P_pTszzamrkUz3YZsXL3bb4mAYIXsPCtKDH1tw,2375
+haystack_experimental/components/summarizers/__init__.py,sha256=BqnfB0ZMb9ufYUjJ4qmmmRLPXa9FT8XKhMWW8G9Zg9Y,221
+haystack_experimental/components/summarizers/llm_summarizer.py,sha256=Rzl3DKWENBKoAiHvgYPsc4ev0WHZGJZj4PBF-FDHiXI,14392
 haystack_experimental/components/writers/__init__.py,sha256=iMdeAaZozza8E6dQ4Lc2uOYIFaR95K7bR9mSeuDqSAA,233
 haystack_experimental/components/writers/chat_message_writer.py,sha256=iu8gmvmRXlqd9S2-9B8p-7C0Y5GTuOI1AqcVKAkrzDc,3502
+haystack_experimental/core/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
+haystack_experimental/core/pipeline/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
+haystack_experimental/core/pipeline/breakpoint.py,sha256=JtwQP8OF5Sdqo0abPRgs1K3SqBkUXhZ53PkeagcK2ZE,5134
+haystack_experimental/dataclasses/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
+haystack_experimental/dataclasses/breakpoints.py,sha256=f0kxYXJRHzk6jAW5Na51MZfUuRIlulhN4oTrGWTpSFE,2095
 haystack_experimental/super_components/__init__.py,sha256=eHD7xrty2PCky_gG3ty19rpM4WfV32TyytM7gJODwl4,110
 haystack_experimental/super_components/indexers/__init__.py,sha256=4VPKnuzVb89Zb4PT6ejYT4s0zJ4I3rwFtcLwsCdQKJA,313
 haystack_experimental/super_components/indexers/sentence_transformers_document_indexer.py,sha256=hfXznLVTgO39xO4GRYgi2Xy-pl4EFKtt13JrGncjvXQ,8519
@@ -30,8 +48,8 @@ haystack_experimental/utils/hallucination_risk_calculator/core_math.py,sha256=8X
 haystack_experimental/utils/hallucination_risk_calculator/dataclasses.py,sha256=3vk9jsbW-7C9n408Qe730qgdXxIOzsTigf4TMLpryvI,2318
 haystack_experimental/utils/hallucination_risk_calculator/openai_planner.py,sha256=-yVQsGzM5rXsAVwolE6sp5W6q1yDw66SiIUuUbPk1ng,11413
 haystack_experimental/utils/hallucination_risk_calculator/skeletonization.py,sha256=qNdBUoFiBjQsI3ovrhd4RyTFmIbv51Goai1Z_l9lG28,5488
-haystack_experimental-0.13.0.dist-info/METADATA,sha256=hb72ME6ftVs_CfpXH8iOathCK0B_lMAQkreiciekaJ0,13033
-haystack_experimental-0.13.0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-haystack_experimental-0.13.0.dist-info/licenses/LICENSE,sha256=93_5nS97uHxptHvK9E8BZgKxLGeIS-rBWT2swIv-X5Y,11368
-haystack_experimental-0.13.0.dist-info/licenses/LICENSE-MIT.txt,sha256=knmLkIKj_6tTrTSVRg9Tq88Kww4UCPLt2I1RGXJv9sQ,1037
-haystack_experimental-0.13.0.dist-info/RECORD,,
+haystack_experimental-0.14.1.dist-info/METADATA,sha256=A1DesX7H05AtVp4S8PD99UIfyFy3cQB_y1a4Oho9Nmc,18566
+haystack_experimental-0.14.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+haystack_experimental-0.14.1.dist-info/licenses/LICENSE,sha256=93_5nS97uHxptHvK9E8BZgKxLGeIS-rBWT2swIv-X5Y,11368
+haystack_experimental-0.14.1.dist-info/licenses/LICENSE-MIT.txt,sha256=knmLkIKj_6tTrTSVRg9Tq88Kww4UCPLt2I1RGXJv9sQ,1037
+haystack_experimental-0.14.1.dist-info/RECORD,,