palimpzest 0.7.21__py3-none-any.whl → 0.8.0__py3-none-any.whl

palimpzest/core/data/context.py

@@ -0,0 +1,393 @@
+from __future__ import annotations
+
+import os
+import re
+from abc import ABC
+from typing import Callable
+
+from pydantic import BaseModel
+from smolagents import CodeAgent, LiteLLMModel
+
+from palimpzest.core.data import context_manager
+from palimpzest.core.data.dataset import Dataset
+from palimpzest.core.lib.schemas import create_schema_from_fields, union_schemas
+from palimpzest.query.operators.logical import ComputeOperator, ContextScan, LogicalOperator, SearchOperator
+from palimpzest.utils.hash_helpers import hash_for_id
+
+PZ_INSTRUCTION = """\n\nYou are a CodeAgent who is a specialist at writing declarative AI programs with the Palimpzest (PZ) library.
+
+Palimpzest is a programming framework which provides you with **semantic operators** (e.g. semantic maps, semantic filters, etc.)
+which are like their traditional counterparts, except they can execute instructions provided in natural language.
+
+For example, if you wanted to write a program to extract the title and abstract from a directory of papers,
+you could write the following in PZ:
+```
+import palimpzest as pz
+from dotenv import load_dotenv
+
+# Load environment variables from .env file
+load_dotenv()
+
+# Define columns for semantic map (sem_map) operation; each column is specified
+# with a dictionary containing the following keys:
+# - "name": the name of the field to compute
+# - "type": the type of the field to compute
+# - "description": the natural language description of the field
+paper_cols = [
+    {"name": "title", "type": str, "description": "the title of the paper"},
+    {"name": "abstract", "type": str, "description": "the paper's abstract"},
+]
+
+# construct the data processing pipeline with PZ
+ds = pz.TextFileDataset(id="papers", path="path/to/papers")
+ds = ds.sem_map(cols)
+
+# optimize and execute the PZ program
+validator = pz.Validator()
+config = pz.QueryProcessorConfig(
+    policy=pz.MaxQuality(),
+    execution_strategy="parallel",
+    max_workers=20,
+    progress=True,
+)
+output = ds.optimize_and_run(config=config, validator=validator)
+
+# write the execution stats to json
+output.execution_stats.to_json("pz_program_stats.json")
+
+# write the output to a CSV and print the output CSV filepath so the user knows where to find it
+output_filepath = "pz_program_output.csv"
+output.to_df().to_csv(output_filepath, index=False)
+print(f"Results at: {output_filepath}")
+```
+
+To initialize a dataset in PZ, simply provide the path to a directory to `pz.TextFileDirectory()`
+(if your data contains text-based files). For example:
+```
+import palimpzest as pz
+from dotenv import load_dotenv
+
+# Load environment variables from .env file
+load_dotenv()
+
+ds = pz.TextFileDataset(id="files", path="path/to/files")
+```
+
+Palimpzest has two primary **semantic operators** which you can use to construct data processing pipelines:
+- sem_filter(predicate: str): executes a semantic filter specified by the natural language predicate on a given PZ dataset
+- sem_map(cols: list[dict]): executes a semantic map to compute the `cols` on a given PZ dataset
+
+As a second example, consider the following PZ program which filters for papers about batteries that are from MIT
+and computes a summary for each one:
+```
+import palimpzest as pz
+from dotenv import load_dotenv
+
+# Load environment variables from .env file
+load_dotenv()
+
+# construct the PZ program
+ds = pz.TextFileDataset(id="papers", path="path/to/research-papers")
+ds = ds.sem_filter("The paper is about batteries")
+ds = ds.sem_filter("The paper is from MIT")
+ds = ds.sem_map([{"name": "summary", "type": str, "description": "A summary of the paper"}])
+
+# optimize and execute the PZ program
+validator = pz.Validator()
+config = pz.QueryProcessorConfig(
+    policy=pz.MaxQuality(),
+    execution_strategy="parallel",
+    max_workers=20,
+    progress=True,
+)
+output = ds.optimize_and_run(config=config, validator=validator)
+
+# write the execution stats to json
+output.execution_stats.to_json("pz_program_stats.json")
+
+# write the output to a CSV and print the output CSV filepath so the user knows where to find it
+output_filepath = "pz_program_output.csv"
+output.to_df().to_csv(output_filepath, index=False)
+print(f"Results at: {output_filepath}")
+```
+
+Be sure to always:
+- execute your program using the `.optimize_and_run()` format shown above
+- call `output.execution_stats.to_json("pz_program_stats.json")` to write execution statistics to disk
+- write your output to CSV and print where you wrote it!
+"""
+
+class Context(Dataset, ABC):
+    """
+    The `Context` class is an abstract base class for root `Datasets` whose data is accessed
+    via user-defined methods. Classes which inherit from this class must implement two methods:
+
+    - `list_filepaths()`: which lists the files that the `Context` has access to.
+    - `read_filepath(path: str)`: which reads the file corresponding to the given `path`.
+
+    A `Context` is a special type of `Dataset` that represents a view over an underlying `Dataset`.
+    Each `Context` has a `name` which uniquely identifies it, as well as a natural language `description`
+    of the data / computation that the `Context` represents. Similar to `Dataset`s, `Context`s can be
+    lazily transformed using functions such as `sem_filter`, `sem_map`, `sem_join`, etc., and they may
+    be materialized or unmaterialized.
+    """
+
+    def __init__(
+            self,
+            id: str,
+            description: str,
+            operator: LogicalOperator,
+            schema: type[BaseModel] | None = None,
+            sources: list[Context] | Context | None = None,
+            materialized: bool = False,
+        ) -> None:
+        """
+        Constructor for the `Context` class.
+
+        Args:
+            id (`str`): a string identifier for the `Context`
+            description (`str`): the description of the data contained within the `Context`
+            operator (`LogicalOperator`): The `LogicalOperator` used to compute this `Context`.
+            schema: (`type[BaseModel] | None`): The schema of this `Context`.
+            sources (`list[Context] | Context | None`): The (list of) `Context(s)` which are input(s) to
+                the operator used to compute this `Context`.
+            materialized (`bool`): True if the `Context` has been computed, False otherwise
+        """
+        # set the description
+        self._description = description
+
+        # set the materialization status
+        self._materialized = materialized
+
+        # compute schema and call parent constructor
+        if schema is None:
+            schema = create_schema_from_fields([{"name": "context", "description": "The context", "type": str}])
+        super().__init__(sources=sources, operator=operator, schema=schema, id=id)
+
+        # set the tools associated with this Context
+        self._tools = [getattr(self, attr) for attr in dir(self) if attr.startswith("tool_")]
+
+        # add Context to ContextManager
+        cm = context_manager.ContextManager()
+        cm.add_context(self)
+
+    @property
+    def description(self) -> str:
+        """The string containing all of the information computed for this `Context`"""
+        return self._description
+
+    @property
+    def materialized(self) -> bool:
+        """The boolean which specifies whether the `Context` has been computed or not"""
+        return self._materialized
+
+    @property
+    def tools(self) -> list[Callable]:
+        """The list of tools associated with this `Context`"""
+        return self._tools
+
+    def __str__(self) -> str:
+        return f"Context(id={self.id}, description={self.description:20s}, materialized={self.materialized})"
+
+    def set_description(self, description: str) -> None:
+        """
+        Update the context's description.
+        """
+        self._description = description
+
+    def set_materialized(self, materialized: str) -> None:
+        """
+        Update the context's materialization status.
+        """
+        self._materialized = materialized
+
+    def compute(self, instruction: str) -> Context:
+        # construct new description and output schema
+        new_id = hash_for_id(instruction)
+        new_description = f"Parent Context ID: {self.id}\n\nThis Context is the result of computing the following instruction on the parent context.\n\nINSTRUCTION: {instruction}\n\n"
+        inter_schema = create_schema_from_fields([{"name": f"result-{new_id}", "desc": "The result from computing the instruction on the input Context",  "type": str}])
+        new_output_schema = union_schemas([self.schema, inter_schema])
+
+        # construct logical operator
+        operator = ComputeOperator(
+            input_schema=self.schema,
+            output_schema=new_output_schema,
+            context_id=new_id,
+            instruction=instruction,
+        )        
+
+        return Context(id=new_id, description=new_description, operator=operator, sources=[self], materialized=False)
+
+    def search(self, search_query: str) -> Context:
+        # construct new description and output schema
+        new_id = hash_for_id(search_query)
+        new_description = f"Parent Context ID: {self.id}\n\nThis Context is the result of searching the parent context for information related to the following query.\n\nSEARCH QUERY: {search_query}\n\n"
+
+        # construct logical operator
+        operator = SearchOperator(
+            input_schema=self.schema,
+            output_schema=self.schema,
+            context_id=new_id,
+            search_query=search_query,
+        )
+
+        return Context(id=new_id, description=new_description, operator=operator, sources=[self], materialized=False)
+
+class TextFileContext(Context):
+    def __init__(self, path: str, id: str, description: str) -> None:
+        """
+        Constructor for the `TextFileContext` class.
+
+        Args:
+            path (str): The path to the file
+            id (str): a string identifier for the `Context`
+            description (str): The description of the data contained within the `Context`
+            kwargs (dict): Keyword arguments containing the `Context's` id and description.
+        """
+        # check that path is a valid file or directory
+        assert os.path.isfile(path) or os.path.isdir(path), f"Path {path} is not a file nor a directory"
+
+        # get list of filepaths
+        self.filepaths = []
+        if os.path.isfile(path):
+            self.filepaths = [path]
+        else:
+            self.filepaths = []
+            for root, _, files in os.walk(path):
+                for file in files:
+                    fp = os.path.join(root, file)
+                    self.filepaths.append(fp)
+            self.filepaths = sorted(self.filepaths)
+
+        # call parent constructor to set id, operator, and schema
+        schema = create_schema_from_fields([{"name": "context", "desc": "The context", "type": str}])
+        super().__init__(
+            id=id,
+            description=description,
+            operator=ContextScan(context=self, output_schema=schema),
+            schema=schema,
+            materialized=True,
+        )
+    def _check_filter_answer_text(self, answer_text: str) -> dict | None:
+        """
+        Return {"passed_operator": True} if and only if "true" is in the answer text.
+        Return {"passed_operator": False} if and only if "false" is in the answer text.
+        Otherwise, return None.
+        """
+        # NOTE: we may be able to eliminate this condition by specifying this JSON output in the prompt;
+        # however, that would also need to coincide with a change to allow the parse_answer_fn to set "passed_operator"
+        if "true" in answer_text.lower():
+            return {"passed_operator": True}
+        elif "false" in answer_text.lower():
+            return {"passed_operator": False}
+        elif "yes" in answer_text.lower():
+            return {"passed_operator": True}
+
+        return None
+
+    def _parse_filter_answer(self, completion_text: str) -> dict[str, list]:
+        """Extract the answer from the completion object for filter operations."""
+        # if the model followed the default instructions, the completion text will place
+        # its answer between "ANSWER:" and "---"
+        regex = re.compile("answer:(.*?)---", re.IGNORECASE | re.DOTALL)
+        matches = regex.findall(completion_text)
+        if len(matches) > 0:
+            answer_text = matches[0].strip()
+            field_answers = self._check_filter_answer_text(answer_text)
+            if field_answers is not None:
+                return field_answers
+
+        # if the first regex didn't find an answer, try taking all the text after "ANSWER:"
+        regex = re.compile("answer:(.*)", re.IGNORECASE | re.DOTALL)
+        matches = regex.findall(completion_text)
+        if len(matches) > 0:
+            answer_text = matches[0].strip()
+            field_answers = self._check_filter_answer_text(answer_text)
+            if field_answers is not None:
+                return field_answers
+
+        # finally, try taking all of the text; throw an exception if this doesn't work
+        field_answers = self._check_filter_answer_text(completion_text)
+        if field_answers is None:
+            raise Exception(f"Could not parse answer from completion text: {completion_text}")
+
+        return field_answers
+
+    # def tool_list_filepaths(self) -> list[str]:
+    #     """
+    #     This tool returns the list of all of the filepaths which the `Context` has access to.
+
+    #     Args:
+    #         None
+        
+    #     Returns:
+    #         list[str]: A list of file paths for all files in the `Context`.
+    #     """
+    #     return self.filepaths
+
+    # def tool_read_filepath(self, path: str) -> str:
+    #     """
+    #     This tool takes a filepath (`path`) as input and returns the content of the file as a string.
+    #     It handles both CSV files and html / regular text files. It does not handle images.
+
+    #     Args:
+    #         path (str): The path to the file to read.
+
+    #     Returns:
+    #         str: The content of the file as a string.
+    #     """
+    #     if path.endswith(".csv"):
+    #         return pd.read_csv(path, encoding="ISO-8859-1").to_string(index=False)
+
+    #     with open(path, encoding='utf-8') as file:
+    #         content = file.read()
+
+    #     return content
+
+    def tool_execute_semantic_operators(self, instruction: str) -> str:
+        """
+        This tool takes an `instruction` as input and invokes an expert to write a semantic data processing pipeline
+        to execute the instruction. The tool returns the path to a CSV file which contains the output of the pipeline.
+
+        For example, the tool could be invoked as follows to extract the title and abstract from a dataset of research papers:
+        ```
+        instruction = "Write a program to extract the title and abstract from each research paper"
+        result_csv_filepath = tool_execute_semantic_operators(instruction)
+        ```
+
+        Args:
+            instruction: The instruction specifying the semantic data processing pipeline that you need to execute.
+
+        Returns:
+            str: the filepath to the CSV containing the output from running the data processing pipeline.
+        """
+        from smolagents import tool
+        @tool
+        def tool_list_filepaths() -> list[str]:
+            """
+            This tool returns the list of all of the filepaths which the `Context` has access to.
+
+            NOTE: You may want to execute this before writing your PZ program to determine where the data lives.
+
+            Args:
+                None
+            
+            Returns:
+                list[str]: A list of file paths for all files in the `Context`.
+            """
+            return self.filepaths
+
+        agent = CodeAgent(
+            model=LiteLLMModel(model_id="openai/o1", api_key=os.getenv("ANTHROPIC_API_KEY")),
+            tools=[tool_list_filepaths],
+            max_steps=20,
+            planning_interval=4,
+            add_base_tools=False,
+            return_full_result=True,
+            additional_authorized_imports=["dotenv", "json", "palimpzest", "pandas"],
+            instructions=PZ_INSTRUCTION,
+        )
+        result = agent.run(instruction)
+        response = result.output
+
+        return response

palimpzest/core/data/context_manager.py

@@ -0,0 +1,163 @@
+from __future__ import annotations
+
+import os
+import pickle
+
+import chromadb
+import chromadb.utils.embedding_functions as embedding_functions
+import tiktoken
+
+from palimpzest.constants import PZ_DIR
+from palimpzest.core.data import context
+
+
+class ContextNotFoundError(Exception):
+    pass
+
+
+class ContextManager:
+    """
+    This class manages the long-term storage of `Contexts`. Each new `Context` is added to
+    the `ContextManager` and serialized to disk. `Contexts` are also indexed, which enables
+    PZ to search for `Context(s)` which may support `search()` and `compute()` operations.
+    """
+    def __init__(self):
+        # create directory with serialized contexts (if it doesn't already exist)
+        self.context_dir = os.path.join(PZ_DIR, "contexts")
+        os.makedirs(self.context_dir, exist_ok=True)
+
+        # create vector store (if it doesn't already exist)
+        self.chroma_dir = os.path.join(PZ_DIR, "chroma")
+        os.makedirs(self.chroma_dir, exist_ok=True)
+        self.chroma_client = chromadb.PersistentClient(self.chroma_dir)
+
+        # pick embedding function based on presence of API key(s)
+        self.emb_fn = None
+        if os.getenv("OPENAI_API_KEY"):
+            self.emb_fn = embedding_functions.OpenAIEmbeddingFunction(
+                api_key=os.getenv("OPENAI_API_KEY"),
+                model_name="text-embedding-3-small"
+            )
+
+        self.index = self.chroma_client.get_or_create_collection("contexts", embedding_function=self.emb_fn)
+
+    @staticmethod
+    def from_pkl(path: str) -> context.Context:
+        """Load a `Context` from its serialized pickle file."""
+        with open(path, "rb") as f:
+            context = pickle.load(f)
+
+        return context
+
+    @staticmethod
+    def to_pkl(context: context.Context, path: str) -> None:
+        """Write the given `Context` to a pickle file at the provided `path`."""
+        with open(path, "wb") as f:
+            pickle.dump(context, f)
+
+    def num_tokens_from_string(self, string: str, encoding_name: str) -> int:
+        """Returns the number of tokens in a text string."""
+        encoding = tiktoken.get_encoding(encoding_name)
+        num_tokens = len(encoding.encode(string))
+        return num_tokens
+
+    def add_context(self, context: context.Context, update: bool = False) -> None:
+        """
+        Add the new `Context` to the `ContextManager` by serializing and writing it to disk.
+
+        Args:
+            context (`Context`): the context to add to the `ContextManager`
+            update (`bool`): whether or not to update an existing context
+
+        TODO: track cost
+        """
+        # return early if the context already exists and we're not performing an update
+        id = context.id
+        context_path = os.path.join(self.context_dir, f"{id}.pkl")
+        if os.path.exists(context_path) and update is False:
+            return
+
+        # write the context to disk
+        ContextManager.to_pkl(context, context_path)
+
+        # compute number of tokens in context.description
+        description = context.description
+        while self.num_tokens_from_string(description, "cl100k_base") > 8192:
+            description = description[:int(0.9*len(description))]
+ 
+        # add context to vector store
+        context_embeddings = self.emb_fn([description])
+        context_payload = {
+            "ids": [context.id],
+            "embeddings": context_embeddings,
+            "metadatas": [{"id": context.id, "materialized": context.materialized}],
+            "documents": [context.description],
+        }
+        if update:
+            self.index.update(**context_payload)
+        else:
+            self.index.add(**context_payload)
+
+    def update_context(self, id: str, description: str, materialized: bool = True) -> None:
+        """
+        Update an existing `Context` with the given `id` to have the given `description`.
+        
+        Args:
+            id (str): the id of the updated `Context`
+            description (str): the update to the description for the specified `Context`
+            materialized (bool): boolean to set the materialization status of the `Context`
+
+        Raises:
+            ContextNotFoundError: if the given `id` doesn't point to a `Context` in the `ContextManger`.
+        """
+        context = self.get_context(id)
+        new_description = context.description + description  # TODO: should description have RESULT replaced on update? as opposed to appending? should description be some pydantic BaseModel?
+        context.set_description(new_description)
+        context.set_materialized(materialized)
+        self.add_context(context, update=True)
+
+    def get_context(self, id: str) -> context.Context:
+        """
+        Returns the `Context` specified by the given `id`.
+
+        Args:
+            id (str): the id of the retrieved `Context`
+
+        Returns:
+            `Context`: the specified `Context`.
+        """
+        context_path = os.path.join(self.context_dir, f"{id}.pkl")
+        try:
+            return ContextManager.from_pkl(context_path)
+        except FileNotFoundError as err:
+            raise ContextNotFoundError from err
+
+    def search_context(self, query: str, k: int = 1, where: dict | None = None) -> list[context.Context]:
+        """
+        Returns the top-k most relevant `Context(s)` for the given query. If provided,
+        the where dictionary will be used to filter the search results.
+
+        TODO:
+        3) update CostModel to account for benefit of using existing Context(s)
+        ---
+        4) unit test
+        5) track cost
+        """
+        # embed the search query
+        query_embeddings = self.emb_fn([query])
+
+        # look up ids of most similar contexts
+        results = self.index.query(
+            query_embeddings=query_embeddings,
+            n_results=k,
+            where=where,
+        )
+        ids = results["ids"][0]
+
+        # load and return Context objects
+        contexts = []
+        for id in ids:
+            context_path = os.path.join(self.context_dir, f"{id}.pkl")
+            contexts.append(ContextManager.from_pkl(context_path))
+
+        return contexts