bump to 0.2.0

agent.py

@@ -6,12 +6,12 @@ from typing import Tuple, List
 
 from omegaconf import OmegaConf
 
-from typing import Optional
 from pydantic import Field, BaseModel
 
 from vectara_agentic.agent import Agent
+from vectara_agentic.agent_config import AgentConfig
 from vectara_agentic.tools import ToolsFactory, VectaraToolFactory
-from vectara_agentic.tools_catalog import summarize_text
+from vectara_agentic.tools_catalog import ToolsCatalog
 
 from dotenv import load_dotenv
 load_dotenv(override=True)
@@ -34,7 +34,7 @@ def extract_components_from_citation(citation: str) -> dict:
 
     return {'volume': int(volume_num), 'reporter': reporter, 'first_page': int(first_page)}
 
-def create_assistant_tools(cfg):
+def create_assistant_tools(cfg, agent_config):
 
     def get_opinion_text(
             case_citation: str = Field(description = citation_description),
@@ -52,6 +52,7 @@ def create_assistant_tools(cfg):
         citation_dict = extract_components_from_citation(case_citation)
         if not citation_dict:
             return f"Citation is invalid: {case_citation}."
+        summarize_text = ToolsCatalog(agent_config).summarize_text
         reporter = citation_dict['reporter']
         volume_num = citation_dict['volume']
         first_page = citation_dict['first_page']
@@ -168,9 +169,8 @@ def create_assistant_tools(cfg):
     class QueryCaselawArgs(BaseModel):
         query: str = Field(..., description="The user query.")
 
-    vec_factory = VectaraToolFactory(vectara_api_key=cfg.api_key, 
-                                     vectara_customer_id=cfg.customer_id, 
-                                     vectara_corpus_id=cfg.corpus_id)
+    vec_factory = VectaraToolFactory(vectara_api_key=cfg.api_key,
+                                     vectara_corpus_key=cfg.corpus_key)
     summarizer = 'vectara-experimental-summary-ext-2023-12-11-med-omni'
 
     ask_caselaw = vec_factory.create_rag_tool(
@@ -188,14 +188,14 @@ def create_assistant_tools(cfg):
                 "diversity_bias": 0.1
             },
             {
-                "type": "udf",
+                "type": "userfn",
                 "user_function": "max(1000 * get('$.score') - hours(seconds(to_unix_timestamp(now()) - to_unix_timestamp(datetime_parse(get('$.document_metadata.decision_date'), 'yyyy-MM-dd')))) / 24 / 365, 0)"
             }
         ],
         n_sentences_before = 2, n_sentences_after = 2, lambda_val = 0.005,
         summary_num_results = 15,
         vectara_summarizer = summarizer,
-        include_citations = False,
+        include_citations = True,
     )
 
     tools_factory = ToolsFactory()
@@ -213,8 +213,6 @@ def create_assistant_tools(cfg):
 
 def get_agent_config() -> OmegaConf:
     cfg = OmegaConf.create({
-        'customer_id': str(os.environ['VECTARA_CUSTOMER_ID']),
-        'corpus_id': str(os.environ['VECTARA_CORPUS_ID']),
         'corpus_key': str(os.environ['VECTARA_CORPUS_KEY']),
         'api_key': str(os.environ['VECTARA_API_KEY']),
         'examples': os.environ.get('QUERY_EXAMPLES', None),
@@ -228,37 +226,39 @@ def initialize_agent(_cfg, agent_progress_callback=None):
     
     legal_assistant_instructions = """
     - You are a helpful legal assistant, with expertise in case law for the state of Alaska.
-    - The ask_caselaw tool is your primary tools for finding information about cases. 
-      Do not use your own knowledge to answer questions.
-    - If the ask_caselaw tool responds that it does not have enough information to answer the query,
-      try to rephrase the query and call the tool again.
-    - If the ask_caselaw tool returns a response along with a ist of references mentioned in its response (in [N] format), 
+    - The ask_caselaw tool is your primary tool for answering user questions. Do not use your own knowledge to answer questions.
+    - If the ask_caselaw tool responds that it does not have enough information to answer the question,
+      try to rephrase the query or break it down into multiple sub-questions, and use the tool again.
+    - If the ask_caselaw tool returns a response along with a list of references mentioned in its response (in [N] format), 
       Format your response to focus on the main response, and use the metadata (such as citations, decision date, or case name) 
       in each relevant reference to provide more context in your response.
+    - If the ask_caselaw tool returns references with case citations, use the `get_case_document_pdf` tool to get a valid URL for the case.
+      and then use that link in your response to provide a clickable link to the case.
     - Citations include 3 components: volume number, reporter, and first page. 
       Here are some examples: '253 P.2d 136', '10 Alaska 11', '6 C.M.A. 3'
     - Never use your internal knowledge to guess citations. Only use citations information provided by a tool or the user.
+    - If you need a valid URL of the full case PDF, call the get_case_document_pdf tool.
+    - If you need a valid URL of the case information page, call the get_case_document_page tool.
+      The text displayed with this URL should be the name_abbreviation of the case (DON'T just say the info can be found here).
+      Don't call the get_case_document_page tool until after you have tried the get_case_document_pdf tool.
+      Don't provide URLs from any other tools. Do not generate URLs yourself.
     - If two cases have conflicting rulings, assume that the case with the more current ruling date is correct.
     - If the response is based on cases that are older than 5 years, make sure to inform the user that the information may be outdated,
       since some case opinions may no longer apply in law.
     - To summarize the case, use the get_opinion_text with summarize=True.
     - Use get_opinion_text with summarize=False only when full text is needed. Consider summarizing the text when possible to make things run faster.
-    - If a user wants to learn more about a case, you can call the get_case_document_pdf tool with the citation to get a valid URL.
-      If this is unsuccessful, call the get_case_document_page tool instead. 
-      The text displayed with this URL should be the name_abbreviation of the case (DON'T just say the info can be found here).
-      Don't call the get_case_document_page tool until after you have tried the get_case_document_pdf tool.
-      Don't provide URLs from any other tools. Do not generate URLs yourself.
-    - When presenting a URL in your response, use the validate_url tool.
+    - When presenting a URL in your response, use the 'validate_url' tool.
     - If a user wants to test their argument, use the ask_caselaw tool to gather information about cases related to their argument 
       and the critique_as_judge tool to determine whether their argument is sound or has issues that must be corrected.
     - Never discuss politics, and always respond politely.
     """
-
+    agent_config = AgentConfig()
     agent = Agent(
-        tools=create_assistant_tools(_cfg),
+        tools=create_assistant_tools(_cfg, agent_config=agent_config),
         topic="Case law in Alaska",
         custom_instructions=legal_assistant_instructions,
         agent_progress_callback=agent_progress_callback,
+        agent_config=agent_config
     )
     agent.report()
     return agent

requirements.txt

@@ -6,4 +6,4 @@ streamlit-feedback==0.1.3
 uuid==1.30
 langdetect==1.0.9
 langcodes==3.4.0
-vectara-agentic==0.1.22
+vectara-agentic==0.2.0

st_app.py

@@ -13,15 +13,6 @@ from agent import initialize_agent, get_agent_config
 
 initial_prompt = "How can I help you today?"
 
-def show_example_questions():        
-    if len(st.session_state.example_messages) > 0 and st.session_state.first_turn:            
-        selected_example = pills("Queries to Try:", st.session_state.example_messages, index=None)
-        if selected_example:
-            st.session_state.ex_prompt = selected_example
-            st.session_state.first_turn = False
-            return True
-    return False
-
 def format_log_msg(log_msg: str):
     max_log_msg_size = 500
     return log_msg if len(log_msg) <= max_log_msg_size else log_msg[:max_log_msg_size]+'...'
@@ -37,7 +28,13 @@ def agent_progress_callback(status_type: AgentStatusType, msg: str):
             latest_message = f"Calling tool {tool_name}..."
         elif status_type == AgentStatusType.TOOL_OUTPUT:
             latest_message = "Analyzing tool output..."
+        elif status_type == AgentStatusType.AGENT_UPDATE:
+            if "Thought:" in msg:
+                latest_message = "Thinking..."
+            else:
+                latest_message = "Updating agent..."
         else:
+            print(f"callback with {status_type} and {msg}")
             return
         
         st.session_state.status.update(label=latest_message)
@@ -46,6 +43,16 @@ def agent_progress_callback(status_type: AgentStatusType, msg: str):
             for log_msg in st.session_state.log_messages:
                 st.markdown(format_log_msg(log_msg), unsafe_allow_html=True)
 
+
+def show_example_questions():        
+    if len(st.session_state.example_messages) > 0 and st.session_state.first_turn:            
+        selected_example = pills("Queries to Try:", st.session_state.example_messages, index=None)
+        if selected_example:
+            st.session_state.ex_prompt = selected_example
+            st.session_state.first_turn = False
+            return True
+    return False
+
 @st.dialog(title="Agent logs", width='large')
 def show_modal():
     for log_msg in st.session_state.log_messages:
@@ -132,8 +139,8 @@ async def launch_bot():
     if st.session_state.prompt:
         with st.chat_message("assistant", avatar='🤖'):
             st.session_state.status = st.status('Processing...', expanded=False)
-            res = st.session_state.agent.chat(st.session_state.prompt)
-            res = escape_dollars_outside_latex(res)
+            response = st.session_state.agent.chat(st.session_state.prompt)
+            res = escape_dollars_outside_latex(response.response)
             message = {"role": "assistant", "content": res, "avatar": '🤖'}
             st.session_state.messages.append(message)
             st.markdown(res)