fabricatio 0.2.6.dev8__tar.gz → 0.2.7.dev1__tar.gz

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/.gitignore

@@ -165,3 +165,4 @@ cython_debug/
 
 extra/scripts/*
 output/
+.vscode/

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fabricatio
-Version: 0.2.6.dev8
+Version: 0.2.7.dev1
 Classifier: License :: OSI Approved :: MIT License
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: 3.12
@@ -24,9 +24,13 @@ Requires-Dist: questionary>=2.1.0
 Requires-Dist: regex>=2024.11.6
 Requires-Dist: rich>=13.9.4
 Requires-Dist: pymilvus>=2.5.4 ; extra == 'rag'
-Requires-Dist: fabricatio[rag] ; extra == 'full'
+Requires-Dist: fabricatio[calc,plot,rag] ; extra == 'full'
+Requires-Dist: sympy>=1.13.3 ; extra == 'calc'
+Requires-Dist: matplotlib>=3.10.1 ; extra == 'plot'
 Provides-Extra: rag
 Provides-Extra: full
+Provides-Extra: calc
+Provides-Extra: plot
 License-File: LICENSE
 Summary: A LLM multi-agent framework.
 Keywords: ai,agents,multi-agent,llm,pyo3

fabricatio-0.2.7.dev1/examples/extract_and_inject/article_rag.py

@@ -0,0 +1,33 @@
+"""Example of proposing a task to a role."""
+
+import asyncio
+
+from fabricatio import Event, Role, Task, WorkFlow, logger
+from fabricatio.actions.rag import RAGTalk
+
+
+async def main() -> None:
+    """Main function."""
+    role = Role(
+        name="Researcher",
+        description="Extract article essence",
+        llm_model="openai/deepseek-r1-distill-llama-70b",
+        llm_rpm=50,
+        llm_tpm=100000,
+        registry={
+            Event.quick_instantiate(e := "answer"): WorkFlow(
+                name="answer",
+                steps=(RAGTalk,),
+            ).update_init_context(collection_name="article_essence"),
+        },
+    )
+
+    task: Task[int] = await role.propose_task(
+        "Answer to all user questions properly and patiently",
+    )
+
+    logger.info(f"Complete {await task.delegate(e)} times qa.")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/examples/extract_and_inject/extract_and_inject.py

@@ -9,6 +9,8 @@ from fabricatio.actions.article import ExtractArticleEssence
 from fabricatio.actions.rag import InjectToDB
 from fabricatio.fs.curd import dump_text, gather_files
 from fabricatio.models.extra import ArticleEssence
+from fabricatio.models.utils import ok
+from pydantic import HttpUrl
 
 
 class SaveToFS(Action):
@@ -31,20 +33,25 @@ async def main() -> None:
     role = Role(
         name="Researcher",
         description="Extract article essence",
-        llm_model="openai/deepseek-r1-distill-llama-70b",
-        llm_rpm=50,
-        llm_tpm=100000,
+        llm_api_endpoint=HttpUrl("https://dashscope.aliyuncs.com/compatible-mode/v1"),
+        llm_model="openai/qwq-plus",
+        llm_stream=True,
+        llm_rpm=500,
+        llm_tpm=5000000,
         registry={
             Event.quick_instantiate("article"): WorkFlow(
                 name="extract",
                 steps=(ExtractArticleEssence(output_key="to_inject"), SaveToFS, InjectToDB(output_key="task_output")),
-                extra_init_context={"collection_name": "article_essence", "output_dir": Path("output")},
+            ).update_init_context(
+                override_inject=True, collection_name="article_essence_max", output_dir=Path("output")
             )
         },
     )
 
-    task: Task[str] = await role.propose_task(
-        "Extract the essence of the article from the files in './bpdf_out'",
+    task: Task[str] = ok(
+        await role.propose_task(
+            "Extract the essence of the article from the files in './bpdf_out'",
+        )
     )
 
     col_name = await task.override_dependencies(gather_files("bpdf_out", "md")).delegate("article")

fabricatio-0.2.7.dev1/examples/write_outline/.gitignore

@@ -0,0 +1,2 @@
+article_briefing.txt
+out.typ

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "fabricatio"
-version = "0.2.6-dev8"
+version = "0.2.7-dev1"
 description = "A LLM multi-agent framework."
 readme = "README.md"
 license = { file = "LICENSE" }
@@ -81,7 +81,13 @@ rag = [
 ]
 
 full = [
-    "fabricatio[rag]",
+    "fabricatio[calc,plot,rag]",
+]
+calc = [
+    "sympy>=1.13.3",
+]
+plot = [
+    "matplotlib>=3.10.1",
 ]
 
 [tool.ruff]
@@ -140,6 +146,9 @@ testpaths = [
 asyncio_mode = "auto"
 asyncio_default_fixture_loop_scope = "function"
 
+[tool.uv.sources]
+fabricatio = { workspace = true }
+
 [[tool.uv.index]]
 url = "https://mirrors.bfsu.edu.cn/pypi/web/simple"
 default = true

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/python/fabricatio/_rust.pyi

@@ -11,7 +11,7 @@ class TemplateManager:
     """
 
     def __init__(
-            self, template_dirs: List[Path], suffix: Optional[str] = None, active_loading: Optional[bool] = None
+        self, template_dirs: List[Path], suffix: Optional[str] = None, active_loading: Optional[bool] = None
     ) -> None:
         """Initialize the template manager.
 
@@ -55,6 +55,16 @@ class TemplateManager:
             RuntimeError: If template rendering fails
         """
 
+    def render_template_raw(self, template: str, data: Dict[str, Any]) -> str:
+        """Render a template with context data.
+
+        Args:
+            template: The template string
+            data: Context dictionary to provide variables to the template
+
+        Returns:
+            Rendered template content as string
+        """
 
 def blake3_hash(content: bytes) -> str:
     """Calculate the BLAKE3 cryptographic hash of data.
@@ -66,7 +76,6 @@ def blake3_hash(content: bytes) -> str:
         Hex-encoded BLAKE3 hash string
     """
 
-
 class BibManager:
     """BibTeX bibliography manager for parsing and querying citation data."""
 

fabricatio-0.2.7.dev1/python/fabricatio/actions/article.py

@@ -0,0 +1,158 @@
+"""Actions for transmitting tasks to targets."""
+
+from pathlib import Path
+from typing import Any, Callable, List, Optional
+
+from fabricatio.fs import safe_text_read
+from fabricatio.journal import logger
+from fabricatio.models.action import Action
+from fabricatio.models.extra import Article, ArticleEssence, ArticleOutline, ArticleProposal
+from fabricatio.models.task import Task
+from fabricatio.models.utils import ok
+
+
+class ExtractArticleEssence(Action):
+    """Extract the essence of article(s) in text format from the paths specified in the task dependencies.
+
+    Notes:
+        This action is designed to extract vital information from articles with Markdown format, which is pure text, and
+        which is converted from pdf files using `magic-pdf` from the `MinerU` project, see https://github.com/opendatalab/MinerU
+    """
+
+    output_key: str = "article_essence"
+    """The key of the output data."""
+
+    async def _execute(
+        self,
+        task_input: Task,
+        reader: Callable[[str], str] = lambda p: Path(p).read_text(encoding="utf-8"),
+        **_,
+    ) -> Optional[List[ArticleEssence]]:
+        if not task_input.dependencies:
+            logger.info(err := "Task not approved, since no dependencies are provided.")
+            raise RuntimeError(err)
+
+        # trim the references
+        contents = ["References".join(c.split("References")[:-1]) for c in map(reader, task_input.dependencies)]
+        return await self.propose(
+            ArticleEssence,
+            contents,
+            system_message=f"# your personal briefing: \n{self.briefing}",
+        )
+
+
+class GenerateArticleProposal(Action):
+    """Generate an outline for the article based on the extracted essence."""
+
+    output_key: str = "article_proposal"
+    """The key of the output data."""
+
+    async def _execute(
+        self,
+        task_input: Optional[Task] = None,
+        article_briefing: Optional[str] = None,
+        article_briefing_path: Optional[str] = None,
+        **_,
+    ) -> Optional[ArticleProposal]:
+        if article_briefing is None and article_briefing_path is None and task_input is None:
+            logger.info("Task not approved, since ")
+            return None
+        if article_briefing_path is None and task_input:
+            article_briefing_path = await self.awhich_pathstr(
+                f"{task_input.briefing}\nExtract the path of file which contains the article briefing."
+            )
+
+        return (
+            await self.propose(
+                ArticleProposal,
+                briefing := (
+                    article_briefing
+                    or safe_text_read(ok(article_briefing_path, "Could not find the path of file to read."))
+                ),
+                **self.prepend_sys_msg(),
+            )
+        ).update_ref(briefing)
+
+
+class GenerateOutline(Action):
+    """Generate the article based on the outline."""
+
+    output_key: str = "article_outline"
+    """The key of the output data."""
+
+    async def _execute(
+        self,
+        article_proposal: ArticleProposal,
+        **_,
+    ) -> Optional[ArticleOutline]:
+        return (
+            await self.propose(
+                ArticleOutline,
+                article_proposal.as_prompt(),
+                **self.prepend_sys_msg(),
+            )
+        ).update_ref(article_proposal)
+
+
+class CorrectProposal(Action):
+    """Correct the proposal of the article."""
+
+    output_key: str = "corrected_proposal"
+
+    async def _execute(self, article_proposal: ArticleProposal, **_) -> Any:
+        return (await self.censor_obj(article_proposal, reference=article_proposal.referenced)).update_ref(
+            article_proposal
+        )
+
+
+class CorrectOutline(Action):
+    """Correct the outline of the article."""
+
+    output_key: str = "corrected_outline"
+    """The key of the output data."""
+
+    async def _execute(
+        self,
+        article_outline: ArticleOutline,
+        **_,
+    ) -> ArticleOutline:
+        return (await self.censor_obj(article_outline, reference=article_outline.referenced.as_prompt())).update_ref(
+            article_outline
+        )
+
+
+class GenerateArticle(Action):
+    """Generate the article based on the outline."""
+
+    output_key: str = "article"
+    """The key of the output data."""
+
+    async def _execute(
+        self,
+        article_outline: ArticleOutline,
+        **_,
+    ) -> Optional[Article]:
+        article: Article = Article.from_outline(article_outline).update_ref(article_outline)
+
+        for c, deps in article.iter_dfs_with_deps():
+            out = await self.correct_obj(
+                c, reference=f"{article_outline.referenced.as_prompt()}\n" + "\n".join(d.display() for d in deps)
+            )
+
+            c.update_from(out)
+        return article
+
+
+class CorrectArticle(Action):
+    """Correct the article based on the outline."""
+
+    output_key: str = "corrected_article"
+    """The key of the output data."""
+
+    async def _execute(
+        self,
+        article: Article,
+        article_outline: ArticleOutline,
+        **_,
+    ) -> Article:
+        return await self.censor_obj(article, reference=article_outline.referenced.as_prompt())

fabricatio-0.2.7.dev1/python/fabricatio/actions/output.py

@@ -0,0 +1,34 @@
+"""Dump the finalized output to a file."""
+
+from pathlib import Path
+from typing import Optional
+
+from fabricatio.models.action import Action
+from fabricatio.models.generic import FinalizedDumpAble
+from fabricatio.models.task import Task
+from fabricatio.models.utils import ok
+
+
+class DumpFinalizedOutput(Action):
+    """Dump the finalized output to a file."""
+
+    output_key: str = "dump_path"
+
+    async def _execute(
+        self,
+        to_dump: FinalizedDumpAble,
+        task_input: Optional[Task] = None,
+        dump_path: Optional[str | Path] = None,
+        **_,
+    ) -> str:
+        dump_path = Path(
+            dump_path
+            or ok(
+                await self.awhich_pathstr(
+                    f"{ok(task_input, 'Neither `task_input` and `dump_path` is provided.').briefing}\n\nExtract a single path of the file, to which I will dump the data."
+                ),
+                "Could not find the path of file to dump the data.",
+            )
+        )
+        ok(to_dump, "Could not dump the data since the path is not specified.").finalized_dump_to(dump_path)
+        return dump_path.as_posix()

fabricatio-0.2.7.dev1/python/fabricatio/actions/rag.py

@@ -0,0 +1,73 @@
+"""Inject data into the database."""
+
+from typing import List, Optional
+
+from fabricatio.capabilities.rag import RAG
+from fabricatio.journal import logger
+from fabricatio.models.action import Action
+from fabricatio.models.generic import PrepareVectorization
+from fabricatio.models.task import Task
+from questionary import text
+
+
+class InjectToDB(Action, RAG):
+    """Inject data into the database."""
+
+    output_key: str = "collection_name"
+
+    async def _execute[T: PrepareVectorization](
+        self, to_inject: Optional[T] | List[Optional[T]], collection_name: str = "my_collection",override_inject:bool=False, **_
+    ) -> Optional[str]:
+        if not isinstance(to_inject, list):
+            to_inject = [to_inject]
+        logger.info(f"Injecting {len(to_inject)} items into the collection '{collection_name}'")
+        if override_inject:
+            self.check_client().client.drop_collection(collection_name)
+        await self.view(collection_name, create=True).consume_string(
+            [
+                t.prepare_vectorization(self.embedding_max_sequence_length)
+                for t in to_inject
+                if isinstance(t, PrepareVectorization)
+            ],
+        )
+
+        return collection_name
+
+
+class RAGTalk(Action, RAG):
+    """RAG-enabled conversational action that processes user questions based on a given task.
+
+    This action establishes an interactive conversation loop where it retrieves context-relevant
+    information to answer user queries according to the assigned task briefing.
+
+    Notes:
+        task_input: Task briefing that guides how to respond to user questions
+        collection_name: Name of the vector collection to use for retrieval (default: "my_collection")
+
+    Returns:
+        Number of conversation turns completed before termination
+    """
+
+    output_key: str = "task_output"
+
+    async def _execute(self, task_input: Task[str], **kwargs) -> int:
+        collection_name = kwargs.get("collection_name", "my_collection")
+        counter = 0
+
+        self.view(collection_name, create=True)
+
+        try:
+            while True:
+                user_say = await text("User: ").ask_async()
+                if user_say is None:
+                    break
+                gpt_say = await self.aask_retrieved(
+                    user_say,
+                    user_say,
+                    extra_system_message=f"You have to answer to user obeying task assigned to you:\n{task_input.briefing}",
+                )
+                print(f"GPT: {gpt_say}")  # noqa: T201
+                counter += 1
+        except KeyboardInterrupt:
+            logger.info(f"executed talk action {counter} times")
+        return counter

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/python/fabricatio/capabilities/correct.py

@@ -10,9 +10,11 @@ from typing import Optional, Unpack, cast
 from fabricatio._rust_instances import TEMPLATE_MANAGER
 from fabricatio.capabilities.review import Review, ReviewResult
 from fabricatio.config import configs
-from fabricatio.models.generic import Display, ProposedAble, WithBriefing
-from fabricatio.models.kwargs_types import CorrectKwargs, ReviewKwargs
+from fabricatio.models.generic import CensoredAble, Display, ProposedAble, WithBriefing
+from fabricatio.models.kwargs_types import CensoredCorrectKwargs, CorrectKwargs, ReviewKwargs
 from fabricatio.models.task import Task
+from questionary import confirm, text
+from rich import print as rprint
 
 
 class Correct(Review):
@@ -55,7 +57,7 @@ class Correct(Review):
         if supervisor_check:
             await review_res.supervisor_check()
         if "default" in kwargs:
-            cast(ReviewKwargs[None], kwargs)["default"] = None
+            cast("ReviewKwargs[None]", kwargs)["default"] = None
         return await self.propose(
             obj.__class__,
             TEMPLATE_MANAGER.render_template(
@@ -89,7 +91,7 @@ class Correct(Review):
             await review_res.supervisor_check()
 
         if "default" in kwargs:
-            cast(ReviewKwargs[None], kwargs)["default"] = None
+            cast("ReviewKwargs[None]", kwargs)["default"] = None
         return await self.ageneric_string(
             TEMPLATE_MANAGER.render_template(
                 configs.templates.correct_template, {"content": input_text, "review": review_res.display()}
@@ -113,3 +115,31 @@ class Correct(Review):
             Optional[Task[T]]: The corrected task, or None if correction fails.
         """
         return await self.correct_obj(task, **kwargs)
+
+    async def censor_obj[M: CensoredAble](
+        self, obj: M, **kwargs: Unpack[CensoredCorrectKwargs[ReviewResult[str]]]
+    ) -> M:
+        """Censor and correct an object based on defined criteria and templates.
+
+        Args:
+            obj (M): The object to be reviewed and corrected.
+            **kwargs (Unpack[CensoredCorrectKwargs]): Additional keyword
+
+        Returns:
+            M: The censored and corrected object.
+        """
+        last_modified_obj = obj
+        modified_obj = None
+        rprint(obj.finalized_dump())
+        while await confirm("Begin to correct obj above with human censorship?").ask_async():
+            while (topic := await text("What is the topic of the obj reviewing?").ask_async()) is not None and topic:
+                ...
+            if (modified_obj := await self.correct_obj(
+                last_modified_obj,
+                topic=topic,
+                **kwargs,
+            )) is None:
+                break
+            last_modified_obj = modified_obj
+            rprint(last_modified_obj.finalized_dump())
+        return modified_obj or last_modified_obj

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/python/fabricatio/capabilities/rag.py

@@ -19,6 +19,7 @@ from fabricatio.models.kwargs_types import (
     EmbeddingKwargs,
     FetchKwargs,
     LLMKwargs,
+    RetrievalKwargs,
 )
 from fabricatio.models.usages import EmbeddingUsage
 from fabricatio.models.utils import MilvusData, ok
@@ -213,6 +214,25 @@ class RAG(EmbeddingUsage):
         self.add_document(await self.pack(text), collection_name or self.safe_target_collection, flush=True)
         return self
 
+    @overload
+    async def afetch_document[V: (int, str, float, bytes)](
+        self,
+        vecs: List[List[float]],
+        desired_fields: List[str],
+        collection_name: Optional[str] = None,
+        similarity_threshold: float = 0.37,
+        result_per_query: int = 10,
+    ) -> List[Dict[str, V]]: ...
+
+    @overload
+    async def afetch_document[V: (int, str, float, bytes)](
+        self,
+        vecs: List[List[float]],
+        desired_fields: str,
+        collection_name: Optional[str] = None,
+        similarity_threshold: float = 0.37,
+        result_per_query: int = 10,
+    ) -> List[V]: ...
     async def afetch_document[V: (int, str, float, bytes)](
         self,
         vecs: List[List[float]],
@@ -275,7 +295,7 @@ class RAG(EmbeddingUsage):
         if isinstance(query, str):
             query = [query]
         return cast(
-            List[str],
+            "List[str]",
             await self.afetch_document(
                 vecs=(await self.vectorize(query)),
                 desired_fields="text",
@@ -283,6 +303,24 @@ class RAG(EmbeddingUsage):
             ),
         )[:final_limit]
 
+    async def aretrieve_compact(
+        self,
+        query: List[str] | str,
+        **kwargs: Unpack[RetrievalKwargs],
+    ) -> str:
+        """Retrieve data from the collection and format it for display.
+
+        Args:
+            query (List[str] | str): The query to be used for retrieval.
+            **kwargs (Unpack[RetrievalKwargs]): Additional keyword arguments for retrieval.
+
+        Returns:
+            str: A formatted string containing the retrieved data.
+        """
+        return TEMPLATE_MANAGER.render_template(
+            configs.templates.retrieved_display_template, {"docs": (await self.aretrieve(query, **kwargs))}
+        )
+
     async def aask_retrieved(
         self,
         question: str,
@@ -313,16 +351,14 @@ class RAG(EmbeddingUsage):
         Returns:
             str: A string response generated after asking with the context of retrieved documents.
         """
-        docs = await self.aretrieve(
+        rendered = await self.aretrieve_compact(
             query or question,
-            final_limit,
+            final_limit=final_limit,
             collection_name=collection_name,
             result_per_query=result_per_query,
             similarity_threshold=similarity_threshold,
         )
 
-        rendered = TEMPLATE_MANAGER.render_template(configs.templates.retrieved_display_template, {"docs": docs[::-1]})
-
         logger.debug(f"Retrieved Documents: \n{rendered}")
         return await self.aask(
             question,

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/python/fabricatio/capabilities/rating.py

@@ -159,7 +159,7 @@ class GiveRating(WithBriefing, LLMUsage):
                 )
             ),
             validator=_validator,
-            **self.prepend(kwargs),
+            **self.prepend_sys_msg(kwargs),
         )
 
     async def draft_rating_criteria(
@@ -191,7 +191,7 @@ class GiveRating(WithBriefing, LLMUsage):
             validator=lambda resp: set(out)
             if (out := JsonCapture.validate_with(resp, list, str, criteria_count)) is not None
             else out,
-            **self.prepend(kwargs),
+            **self.prepend_sys_msg(kwargs),
         )
 
     async def draft_rating_criteria_from_examples(
@@ -244,7 +244,7 @@ class GiveRating(WithBriefing, LLMUsage):
                 validator=lambda resp: JsonCapture.validate_with(
                     resp, target_type=list, elements_type=str, length=reasons_count
                 ),
-                **self.prepend(kwargs),
+                **self.prepend_sys_msg(kwargs),
             )
         )
         # extract certain mount of criteria from reasons according to their importance and frequency
@@ -301,7 +301,7 @@ class GiveRating(WithBriefing, LLMUsage):
                 for pair in windows
             ],
             validator=lambda resp: JsonCapture.validate_with(resp, target_type=float),
-            **self.prepend(kwargs),
+            **self.prepend_sys_msg(kwargs),
         )
         weights = [1]
         for rw in relative_weights:

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/python/fabricatio/capabilities/review.py

@@ -121,7 +121,7 @@ class ReviewResult[T](ProposedAble, Display):
             ReviewResult[K]: The current instance with updated reference type.
         """
         self._ref = ref  # pyright: ignore [reportAttributeAccessIssue]
-        return cast(ReviewResult[K], self)
+        return cast("ReviewResult[K]", self)
 
     def deref(self) -> T:
         """Retrieve the referenced object that was reviewed.
@@ -200,7 +200,7 @@ class Review(GiveRating, Propose):
             ReviewResult[Task[T]]: A review result containing identified problems and proposed solutions,
                 with a reference to the original task.
         """
-        return cast(ReviewResult[Task[T]], await self.review_obj(task, **kwargs))
+        return cast('ReviewResult[Task[T]]', await self.review_obj(task, **kwargs))
 
     async def review_string(
         self,

{fabricatio-0.2.6.dev8 → fabricatio-0.2.7.dev1}/python/fabricatio/capabilities/task.py

@@ -23,7 +23,7 @@ class ProposeTask(WithBriefing, Propose):
         self,
         prompt: str,
         **kwargs: Unpack[ValidateKwargs[Task[T]]],
-    ) -> Task[T]:
+    ) -> Optional[Task[T]]:
         """Asynchronously proposes a task based on a given prompt and parameters.
 
         Parameters:
@@ -37,7 +37,7 @@ class ProposeTask(WithBriefing, Propose):
             logger.error(err := f"{self.name}: Prompt must be provided.")
             raise ValueError(err)
 
-        return await self.propose(Task, prompt, **self.prepend(cast(Dict[str, Any], kwargs)))
+        return await self.propose(Task, prompt, **self.prepend_sys_msg(cast("Dict[str, Any]", kwargs)))
 
 
 class HandleTask(WithBriefing, ToolBoxUsage):
@@ -81,7 +81,7 @@ class HandleTask(WithBriefing, ToolBoxUsage):
         return await self.aask_validate(
             question=q,
             validator=_validator,
-            **self.prepend(cast(Dict[str, Any], kwargs)),
+            **self.prepend_sys_msg(cast("Dict[str, Any]", kwargs)),
         )
 
     async def handle_fine_grind(