janus-llm 3.4.0__tar.gz → 3.4.2__tar.gz

{janus_llm-3.4.0 → janus_llm-3.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: janus-llm
-Version: 3.4.0
+Version: 3.4.2
 Summary: A transcoding library using LLMs.
 Home-page: https://github.com/janus-llm/janus-llm
 License: Apache 2.0

{janus_llm-3.4.0 → janus_llm-3.4.2}/janus/__init__.py

@@ -5,7 +5,7 @@ from langchain_core._api.deprecation import LangChainDeprecationWarning
 from janus.converter.translate import Translator
 from janus.metrics import *  # noqa: F403
 
-__version__ = "3.4.0"
+__version__ = "3.4.2"
 
 # Ignoring a deprecation warning from langchain_core that I can't seem to hunt down
 warnings.filterwarnings("ignore", category=LangChainDeprecationWarning)

{janus_llm-3.4.0 → janus_llm-3.4.2}/janus/cli.py

@@ -248,6 +248,13 @@ def translate(
             "If unspecificed, model's default max will be used.",
         ),
     ] = None,
+    skip_refiner: Annotated[
+        bool,
+        typer.Option(
+            "--skip-refiner",
+            help="Whether to skip the refiner for generating output",
+        ),
+    ] = True,
 ):
     try:
         target_language, target_version = target_lang.split("-")
@@ -274,6 +281,7 @@ def translate(
         db_config=collections_config,
         splitter_type=splitter_type,
         skip_context=skip_context,
+        skip_refiner=skip_refiner,
     )
     translator.translate(input_dir, output_dir, overwrite, collection)
 
@@ -395,6 +403,13 @@ def document(
             "If unspecificed, model's default max will be used.",
         ),
     ] = None,
+    skip_refiner: Annotated[
+        bool,
+        typer.Option(
+            "--skip-refiner",
+            help="Whether to skip the refiner for generating output",
+        ),
+    ] = True,
 ):
     model_arguments = dict(temperature=temperature)
     collections_config = get_collections_config()
@@ -407,6 +422,7 @@ def document(
         db_path=db_loc,
         db_config=collections_config,
         splitter_type=splitter_type,
+        skip_refiner=skip_refiner,
         skip_context=skip_context,
     )
     if doc_mode == "madlibs":
@@ -520,6 +536,13 @@ def diagram(
             click_type=click.Choice(list(CUSTOM_SPLITTERS.keys())),
         ),
     ] = "file",
+    skip_refiner: Annotated[
+        bool,
+        typer.Option(
+            "--skip-refiner",
+            help="Whether to skip the refiner for generating output",
+        ),
+    ] = True,
 ):
     model_arguments = dict(temperature=temperature)
     collections_config = get_collections_config()
@@ -533,6 +556,7 @@ def diagram(
         diagram_type=diagram_type,
         add_documentation=add_documentation,
         splitter_type=splitter_type,
+        skip_refiner=skip_refiner,
         skip_context=skip_context,
     )
     diagram_generator.translate(input_dir, output_dir, overwrite, collection)

{janus_llm-3.4.0 → janus_llm-3.4.2}/janus/converter/converter.py

@@ -1,6 +1,5 @@
 import functools
 import json
-import math
 import time
 from pathlib import Path
 from typing import Any, List, Optional, Tuple
@@ -77,6 +76,7 @@ class Converter:
         prune_node_types: tuple[str, ...] = (),
         splitter_type: str = "file",
         refiner_type: str = "basic",
+        skip_refiner: bool = True,
         skip_context: bool = False,
     ) -> None:
         """Initialize a Converter instance.
@@ -98,6 +98,8 @@ class Converter:
             splitter_type: The type of splitter to use. Valid values are `"file"`,
                 `"tag"`, `"chunk"`, `"ast-strict"`, and `"ast-flex"`.
             refiner_type: The type of refiner to use. Valid values are `"basic"`.
+            skip_refiner: Whether to skip the refiner.
+            skip_context: Whether to skip adding context to the prompt.
         """
         self._changed_attrs: set = set()
 
@@ -133,6 +135,8 @@ class Converter:
         self._refiner_type: str
         self._refiner: Refiner
 
+        self.skip_refiner = skip_refiner
+
         self.set_splitter(splitter_type=splitter_type)
         self.set_refiner(refiner_type=refiner_type)
         self.set_model(model_name=model, **model_arguments)
@@ -293,7 +297,7 @@ class Converter:
         """
         if self._refiner_type == "basic":
             self._refiner = BasicRefiner(
-                "basic_refinement", self._model_name, self._source_language
+                "basic_refinement", self._model_id, self._source_language
             )
         else:
             raise ValueError(f"Error: unknown refiner type {self._refiner_type}")
@@ -598,40 +602,41 @@ class Converter:
         self._parser.set_reference(block.original)
 
         # Retries with just the output and the error
-        n1 = round(self.max_prompts ** (1 / 3))
+        n1 = round(self.max_prompts ** (1 / 2))
 
         # Retries with the input, output, and error
-        n2 = round((self.max_prompts // n1) ** (1 / 2))
+        n2 = round(self.max_prompts // n1)
 
         # Retries with just the input
-        n3 = math.ceil(self.max_prompts / (n1 * n2))
-        # Make replacements in the prompt
         if not self.skip_context:
             self._make_prompt_additions(block)
-
-        refine_output = RefinerParser(
-            parser=self._parser,
-            initial_prompt=self._prompt.format(**{"SOURCE_CODE": block.original.text}),
-            refiner=self._refiner,
-            max_retries=n1,
-            llm=self._llm,
-        )
-        retry = RetryWithErrorOutputParser.from_llm(
-            llm=self._llm,
-            parser=refine_output,
-            max_retries=n2,
-        )
+        if not self.skip_refiner:  # Make replacements in the prompt
+            refine_output = RefinerParser(
+                parser=self._parser,
+                initial_prompt=self._prompt.format(
+                    **{"SOURCE_CODE": block.original.text}
+                ),
+                refiner=self._refiner,
+                max_retries=n1,
+                llm=self._llm,
+            )
+        else:
+            refine_output = RetryWithErrorOutputParser.from_llm(
+                llm=self._llm,
+                parser=self._parser,
+                max_retries=n1,
+            )
         completion_chain = self._prompt | self._llm
         chain = RunnableParallel(
             completion=completion_chain, prompt_value=self._prompt
-        ) | RunnableLambda(lambda x: retry.parse_with_prompt(**x))
-        for _ in range(n3):
+        ) | RunnableLambda(lambda x: refine_output.parse_with_prompt(**x))
+        for _ in range(n2):
             try:
                 return chain.invoke({"SOURCE_CODE": block.original.text})
             except OutputParserException:
                 pass
 
-        raise OutputParserException(f"Failed to parse after {n1*n2*n3} retries")
+        raise OutputParserException(f"Failed to parse after {n1*n2} retries")
 
     def _get_output_obj(
         self, block: TranslatedCodeBlock
@@ -672,6 +677,10 @@ class Converter:
             ]
         )
 
+        if not hasattr(self._prompt, "messages"):
+            log.debug("Skipping additions to prompt, no messages found on prompt object!")
+            return
+
         # Iterate through existing messages to find and update the system message
         for i, message in enumerate(self._prompt.messages):
             if isinstance(message, SystemMessagePromptTemplate):

{janus_llm-3.4.0 → janus_llm-3.4.2}/janus/parsers/refiner_parser.py

@@ -40,7 +40,9 @@ class RefinerParser(BaseOutputParser):
                 return self.parser.parse(text)
             except OutputParserException as oe:
                 err = str(oe)
-                new_prompt, prompt_arguments = self.refiner.refine(last_prompt, text, err)
+                new_prompt, prompt_arguments = self.refiner.refine(
+                    self.initial_prompt, last_prompt, text, err
+                )
                 new_chain = new_prompt | self.llm
                 text = new_chain.invoke(prompt_arguments)
                 last_prompt = new_prompt.format(**prompt_arguments)

{janus_llm-3.4.0 → janus_llm-3.4.2}/janus/refiners/refiner.py

@@ -5,7 +5,12 @@ from janus.llm.models_info import MODEL_PROMPT_ENGINES
 
 class Refiner:
     def refine(
-        self, original_prompt: str, original_output: str, errors: str, **kwargs
+        self,
+        original_prompt: str,
+        previous_prompt: str,
+        previous_output: str,
+        errors: str,
+        **kwargs,
     ) -> tuple[ChatPromptTemplate, dict[str, str]]:
         """Creates a new prompt based on feedback from original results
 
@@ -24,22 +29,27 @@ class BasicRefiner(Refiner):
     def __init__(
         self,
         prompt_name: str,
-        model_name: str,
+        model_id: str,
         source_language: str,
     ) -> None:
         """Basic refiner, asks llm to fix output of previous prompt given errors
 
         Arguments:
             prompt_name: refinement prompt name to use
-            model_name: name of llm to use
+            model_id: ID of the llm to use. Found in models_info.py
             source_language: source_langauge to use
         """
         self._prompt_name = prompt_name
-        self._model_name = model_name
+        self._model_id = model_id
         self._source_language = source_language
 
     def refine(
-        self, original_prompt: str, original_output: str, errors: str, **kwargs
+        self,
+        original_prompt: str,
+        previous_prompt: str,
+        previous_output: str,
+        errors: str,
+        **kwargs,
     ) -> tuple[ChatPromptTemplate, dict[str, str]]:
         """Creates a new prompt based on feedback from original results
 
@@ -51,13 +61,13 @@ class BasicRefiner(Refiner):
         Returns:
             Tuple of new prompt and prompt arguments
         """
-        prompt_engine = MODEL_PROMPT_ENGINES[self._model_name](
+        prompt_engine = MODEL_PROMPT_ENGINES[self._model_id](
             prompt_template=self._prompt_name,
             source_language=self._source_language,
         )
         prompt_arguments = {
             "ORIGINAL_PROMPT": original_prompt,
-            "OUTPUT": original_output,
+            "OUTPUT": previous_output,
             "ERRORS": errors,
         }
         return prompt_engine.prompt, prompt_arguments

{janus_llm-3.4.0 → janus_llm-3.4.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "janus-llm"
-version = "3.4.0"
+version = "3.4.2"
 description = "A transcoding library using LLMs."
 authors = ["Michael Doyle <mdoyle@mitre.org>", "Chris Glasz <cglasz@mitre.org>",
            "Chris Tohline <ctohline@mitre.org>", "William Macke <wmacke@mitre.org>",