vectara-agentic 0.3.0__tar.gz → 0.3.1__tar.gz

{vectara_agentic-0.3.0/vectara_agentic.egg-info → vectara_agentic-0.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vectara_agentic
-Version: 0.3.0
+Version: 0.3.1
 Summary: A Python package for creating AI Assistants and AI Agents with Vectara
 Home-page: https://github.com/vectara/py-vectara-agentic
 Author: Ofer Mendelevitch
@@ -61,6 +61,7 @@ Requires-Dist: python-dotenv==1.0.1
 Requires-Dist: tiktoken==0.9.0
 Requires-Dist: cloudpickle>=3.1.1
 Requires-Dist: httpx==0.28.1
+Requires-Dist: commonmark==0.9.1
 Dynamic: author
 Dynamic: author-email
 Dynamic: classifier

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1}/requirements.txt

@@ -43,3 +43,4 @@ python-dotenv==1.0.1
 tiktoken==0.9.0
 cloudpickle>=3.1.1
 httpx==0.28.1
+commonmark==0.9.1

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1}/vectara_agentic/_prompts.py

@@ -28,11 +28,11 @@ GENERAL_INSTRUCTIONS = """
 - If after retrying you can't get the information or answer the question, respond with "I don't know".
 - Handling references and citations:
   1) Include references and citations in your response to increase the credibility of your answer.
-  2) Citations should be included in the response, along with URLs, as in-text markers, such as [1](https://www.xxx.com), [2](https://www.yyy.com), etc.
+  2) Citations should be included in the response, along with URLs, as in-text markers, such as [1](https://www.xxx.com), [2](https://www.yyy.com/doc.pdf#page=2), etc.
      You can also replace the number with a word or sentence that describes the reference, such as "[according to Nvidia 10-K](https://www.xxx.com)".
      When adding a citation inline in the text, make sure to use proper spacing and punctuation.
   3) If a URL is a PDF file, and the tool also provided a page number - then combine the URL and page number in your response.
-     For example, if the URL returned from the tool is "https://www.xxx.com/doc.pdf" and "page=5", then the combined URL would be "https://www.xxx.com/doc.pdf#page=5".
+     For example, if the URL returned from the tool is "https://www.xxx.com/doc.pdf" and "page='5'", then the combined URL would be "https://www.xxx.com/doc.pdf#page=5".
   4) Where possible, integrate citations into the text of your response, such as "According to the [Nvidia 10-K](https://www.xxx.com), the revenue in 2021 was $10B".
   5) Only include citations if provided with a valid URL as part of the tool's output (directly or in the metadata).
   6) If a tool returns in the metadata invalid URLs or an empty URL (e.g. "[[1]()]"), ignore it and do not include that citation or reference in your response.

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1}/vectara_agentic/_version.py

@@ -1,4 +1,4 @@
 """
 Define the version of the package.
 """
-__version__ = "0.3.0"
+__version__ = "0.3.1"

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1}/vectara_agentic/hhem.py

@@ -1,6 +1,34 @@
 """Vectara HHEM (Hypothesis Hypothetical Evaluation Model) client."""
 
 import requests
+from commonmark import Parser
+
+
+def markdown_to_text(md: str) -> str:
+    """
+    Convert a Markdown-formatted string into plain text.
+    """
+    parser = Parser()
+    ast = parser.parse(md)
+    out: list[str] = []
+
+    def recurse(node):
+        if node.t in ("text", "code", "html_inline"):
+            out.append(node.literal or "")
+        elif node.t == "softbreak":
+            out.append(" ")
+        elif node.t == "linebreak":
+            out.append("\n")
+        child = getattr(node, "first_child", None)
+        while child is not None:
+            recurse(child)
+            child = getattr(child, "next", None)
+
+    recurse(ast)
+    text = "".join(out)
+    # collapse runs of spaces but preserve newlines
+    lines = [" ".join(line.split()) for line in text.splitlines()]
+    return "\n".join(line if line.strip() else "" for line in lines)
 
 
 class HHEM:
@@ -23,9 +51,18 @@ class HHEM:
         Raises:
             requests.exceptions.RequestException: If there is a network-related error or the API call fails.
         """
+
+        # clean response from any markdown or other formatting.
+        try:
+            clean_hypothesis = markdown_to_text(hypothesis)
+        except Exception as e:
+            # If markdown parsing fails, use the original text
+            raise ValueError(f"Markdown parsing of hypothesis failed: {e}") from e
+
+        # compute HHEM with Vectara endpoint
         payload = {
             "model_parameters": {"model_name": "hhem_v2.3"},
-            "generated_text": hypothesis,
+            "generated_text": clean_hypothesis,
             "source_texts": [context],
         }
         headers = {

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1}/vectara_agentic/tools.py

@@ -274,22 +274,17 @@ class VectaraToolFactory:
                 for i, result in enumerate(results, 1):
                     result_str = f"**Result #{i}**\n"
                     result_str += f"Document ID: {result['metadata']['document_id']}\n"
-                    result_str += (
-                        f"Matches: {len(result['metadata']['matching_text'])}\n"
-                    )
-
                     if summarize and result["text"]:
                         result_str += f"Summary: {result['text']}\n"
 
-                    # Add sample matching text if available
-                    if result["metadata"]["matching_text"]:
-                        sample_matches = result["metadata"]["matching_text"][
-                            :2
-                        ]  # Show first 2 matches
-                        result_str += f"Sample matches: {', '.join(sample_matches)}\n"
-
+                    # Add all matching text if available
+                    matches = result["metadata"]["matching_text"]
+                    if matches:
+                        result_str += ''.join(
+                            f"Match #{inx} Text: {match}\n"
+                            for inx, match in enumerate(matches, 1)
+                        )
                     formatted_results.append(result_str)
-
                 return "\n".join(formatted_results)
 
             return create_human_readable_output(res, format_search_results)
@@ -448,6 +443,7 @@ class VectaraToolFactory:
             vectara_base_url=vectara_base_url,
             vectara_verify_ssl=vectara_verify_ssl,
         )
+        keys_to_ignore = ["lang", "offset", "len"]
 
         # Dynamically generate the RAG function
         def rag_function(*args: Any, **kwargs: Any) -> dict:
@@ -527,7 +523,6 @@ class VectaraToolFactory:
             matches = re.findall(pattern, response.response)
             citation_numbers = sorted(set(int(match) for match in matches))
             citation_metadata = {}
-            keys_to_ignore = ["lang", "offset", "len"]
             for citation_number in citation_numbers:
                 metadata = {
                     k: v
@@ -549,21 +544,30 @@ class VectaraToolFactory:
                     }
             if fcs:
                 citation_metadata["fcs"] = fcs
-
             res = {"text": response.response, "metadata": citation_metadata}
 
             # Create human-readable output with citation formatting
             def format_rag_response(result):
                 text = result["text"]
-                metadata = result["metadata"]
 
                 # Format citations if present
+                metadata = result["metadata"]
                 citation_info = []
                 for key, value in metadata.items():
                     if key.isdigit():
-                        url = value.get("document", {}).get("url", None)
-                        if url:
-                            citation_info.append(f"[{key}]: {url}")
+                        doc = value.get("document", {})
+                        doc_metadata = f"{key}: " + "; ".join(
+                            [
+                                f"{k}='{v}'"
+                                for k, v in doc.items()
+                            ] +
+                            [
+                                f"{k}='{v}'"
+                                for k, v in value.items()
+                                if k not in ["document"] + keys_to_ignore
+                            ]
+                        )
+                        citation_info.append(doc_metadata)
                 if citation_info:
                     text += "\n\nCitations:\n" + "\n".join(citation_info)
 

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1/vectara_agentic.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vectara_agentic
-Version: 0.3.0
+Version: 0.3.1
 Summary: A Python package for creating AI Assistants and AI Agents with Vectara
 Home-page: https://github.com/vectara/py-vectara-agentic
 Author: Ofer Mendelevitch
@@ -61,6 +61,7 @@ Requires-Dist: python-dotenv==1.0.1
 Requires-Dist: tiktoken==0.9.0
 Requires-Dist: cloudpickle>=3.1.1
 Requires-Dist: httpx==0.28.1
+Requires-Dist: commonmark==0.9.1
 Dynamic: author
 Dynamic: author-email
 Dynamic: classifier

{vectara_agentic-0.3.0 → vectara_agentic-0.3.1}/vectara_agentic.egg-info/requires.txt

@@ -43,3 +43,4 @@ python-dotenv==1.0.1
 tiktoken==0.9.0
 cloudpickle>=3.1.1
 httpx==0.28.1
+commonmark==0.9.1