langroid 0.1.249__tar.gz → 0.1.250__tar.gz

{langroid-0.1.249 → langroid-0.1.250}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.1.249
+Version: 0.1.250
 Summary: Harness LLMs with Multi-Agent Programming
 License: MIT
 Author: Prasad Chalasani
@@ -237,6 +237,8 @@ teacher_task.run()
 <summary> <b>Click to expand</b></summary>
 
 - **May 2024:** 
+  - [Much-Improved Citation](https://github.com/langroid/langroid/issues/477) 
+    generation and display when using `DocChatAgent`.
   - `gpt-4o` is now the default LLM throughout; Update tests and examples to work 
     with this LLM; use tokenizer corresponding to the LLM.
   - `gemini 1.5 pro` support via `litellm`

{langroid-0.1.249 → langroid-0.1.250}/README.md

@@ -123,6 +123,8 @@ teacher_task.run()
 <summary> <b>Click to expand</b></summary>
 
 - **May 2024:** 
+  - [Much-Improved Citation](https://github.com/langroid/langroid/issues/477) 
+    generation and display when using `DocChatAgent`.
   - `gpt-4o` is now the default LLM throughout; Update tests and examples to work 
     with this LLM; use tokenizer corresponding to the LLM.
   - `gemini 1.5 pro` support via `litellm`

{langroid-0.1.249 → langroid-0.1.250}/langroid/agent/special/doc_chat_agent.py

@@ -14,7 +14,6 @@ pip install "langroid[hf-embeddings]"
 """
 
 import logging
-import re
 from functools import cache
 from typing import Any, Dict, List, Optional, Set, Tuple, no_type_check
 
@@ -31,6 +30,7 @@ from langroid.agent.special.relevance_extractor_agent import (
     RelevanceExtractorAgentConfig,
 )
 from langroid.agent.task import Task
+from langroid.agent.tools.retrieval_tool import RetrievalTool
 from langroid.embedding_models.models import OpenAIEmbeddingsConfig
 from langroid.language_models.base import StreamingIfAllowed
 from langroid.language_models.openai_gpt import OpenAIChatModel, OpenAIGPTConfig
@@ -82,21 +82,47 @@ except ImportError:
     pass
 
 
-def extract_citations(text: str) -> List[int]:
-    # Find all patterns that match [[<numbers>]]
-    matches = re.findall(r"\[\[([\d,]+)\]\]", text)
+def extract_markdown_references(md_string: str) -> list[int]:
+    """
+    Extracts markdown references (e.g., [^1], [^2]) from a string and returns
+    them as a sorted list of integers.
+
+    Args:
+        md_string (str): The markdown string containing references.
+
+    Returns:
+        list[int]: A sorted list of unique integers from the markdown references.
+    """
+    import re
+
+    # Regex to find all occurrences of [^<number>]
+    matches = re.findall(r"\[\^(\d+)\]", md_string)
+    # Convert matches to integers, remove duplicates with set, and sort
+    return sorted(set(int(match) for match in matches))
 
-    # Initialize a set to hold distinct citation numbers
-    citations: Set[int] = set()
 
-    # Process each match
-    for match in matches:
-        # Split numbers by comma and convert to integers
-        numbers = match.split(",")
-        citations.update(int(number) for number in numbers)
+def format_footnote_text(content: str, width: int = 80) -> str:
+    """
+    Formats the content part of a footnote (i.e. not the first line that
+    appears right after the reference [^4])
+    It wraps the text so that no line is longer than the specified width and indents
+    lines as necessary for markdown footnotes.
+
+    Args:
+        content (str): The text of the footnote to be formatted.
+        width (int): Maximum width of the text lines.
 
-    # Return a sorted list of unique citations
-    return sorted(citations)
+    Returns:
+        str: Properly formatted markdown footnote text.
+    """
+    import textwrap
+
+    # Wrap the text to the specified width
+    wrapped_lines = textwrap.wrap(content, width)
+    if len(wrapped_lines) == 0:
+        return ""
+    indent = "    "  # Indentation for markdown footnotes
+    return indent + ("\n" + indent).join(wrapped_lines)
 
 
 class DocChatAgentConfig(ChatAgentConfig):
@@ -438,6 +464,13 @@ class DocChatAgent(ChatAgent):
         self.setup_documents(docs, filter=self.config.filter)
         return len(docs)
 
+    def retrieval_tool(self, msg: RetrievalTool) -> str:
+        """Handle the RetrievalTool message"""
+        self.config.retrieve_only = True
+        self.config.parsing.n_similar_docs = msg.num_results
+        content_doc = self.answer_from_docs(msg.query)
+        return content_doc.content
+
     @staticmethod
     def document_compatible_dataframe(
         df: pd.DataFrame,
@@ -808,14 +841,15 @@ class DocChatAgent(ChatAgent):
         final_answer = answer_doc.content.strip()
         show_if_debug(final_answer, "SUMMARIZE_RESPONSE= ")
 
-        citations = extract_citations(final_answer)
+        citations = extract_markdown_references(final_answer)
 
         citations_str = ""
         if len(citations) > 0:
             # append [i] source, content for each citation
             citations_str = "\n".join(
                 [
-                    f"[{c}] {passages[c-1].metadata.source}\n{passages[c-1].content}"
+                    f"[^{c}] {passages[c-1].metadata.source}"
+                    f"\n{format_footnote_text(passages[c-1].content)}"
                     for c in citations
                 ]
             )

langroid-0.1.250/langroid/agent/tools/retrieval_tool.py

@@ -0,0 +1,29 @@
+from typing import List
+
+from langroid.agent.tool_message import ToolMessage
+
+
+class RetrievalTool(ToolMessage):
+    """Retrieval tool, only to be used by a DocChatAgent."""
+
+    request: str = "retrieval_tool"
+    purpose: str = """
+            To retrieve up to <num_results> passages from a document-set, that are 
+            relevant to a <query>, which could be a question or simply a topic or 
+            search phrase. 
+            """
+    query: str
+    num_results: int
+
+    @classmethod
+    def examples(cls) -> List["ToolMessage"]:
+        return [
+            cls(
+                query="What are the eligibility criteria for the scholarship?",
+                num_results=3,
+            ),
+            cls(
+                query="Self-Attention mechanism in RNNs",
+                num_results=5,
+            ),
+        ]

{langroid-0.1.249 → langroid-0.1.250}/langroid/prompts/templates.py

@@ -55,13 +55,14 @@ SUMMARY_ANSWER_PROMPT_GPT4 = f"""
         information in these extracts, even if your answer is factually incorrect, 
         and even if the answer contradicts other parts of the document. The only 
         important thing is that your answer is consistent with and supported by the 
-        extracts. Compose your complete answer, inserting CITATIONS 
-        in the format [[i,j,...]] where i,j,... are the extract NUMBERS you are citing.
+        extracts. Compose your complete answer, inserting CITATIONS in MARKDOWN format
+        [^i][^j] where i,j,... are the extract NUMBERS you are 
+        citing.
         For example your answer might look like this (NOTE HOW multiple citations
-        are grouped as [[2,5]]):
+        are grouped as [^2][^5]):
         
-        Beethoven composed the 9th symphony in 1824. [[1]] After that he became deaf
-        and could not hear his own music. [[2,5]] He was a prolific composer and
+        Beethoven composed the 9th symphony in 1824.[^1] After that he became deaf
+        and could not hear his own music. [^2][^5]. He was a prolific composer and
         wrote many famous pieces.
         
         NUMBERED EXTRACTS:

{langroid-0.1.249 → langroid-0.1.250}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.1.249"
+version = "0.1.250"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"