updated

Dockerfile

@@ -14,6 +14,9 @@ RUN useradd -m -u 1000 user
 USER user
 ENV HOME /home/user
 ENV PATH $HOME/.local/bin:$PATH
+ENV TIKTOKEN_CACHE_DIR $HOME/.cache/tiktoken
+
+RUN mkdir -p $HOME/.cache/tiktoken
 
 WORKDIR $HOME
 RUN mkdir app

agent.py

@@ -2,7 +2,7 @@ import os
 import re
 import requests
 import json
-from typing import Tuple, List
+from typing import Tuple, List, Optional
 
 from omegaconf import OmegaConf
 
@@ -12,6 +12,7 @@ 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 ToolsCatalog
+from vectara_agentic.types import ModelProvider, AgentType
 
 from dotenv import load_dotenv
 load_dotenv(override=True)
@@ -39,9 +40,10 @@ class AgentTools:
         self.tools_factory = ToolsFactory()
         self.agent_config = agent_config
         self.cfg = _cfg
-        self.vec_factory = VectaraToolFactory(vectara_api_key=_cfg.api_key,
-                                              vectara_corpus_key=_cfg.corpus_key)
-
+        self.vec_factory = VectaraToolFactory(
+            vectara_api_key=_cfg.api_key,
+            vectara_corpus_key=_cfg.corpus_key,
+        )
 
     def get_opinion_text(
             self,
@@ -212,15 +214,13 @@ class AgentTools:
 
     def get_tools(self):
         class QueryCaselawArgs(BaseModel):
-            query: str = Field(..., description="The user query.")
+            citations: Optional[str] = Field(description = citation_description, default=None)
 
-        vec_factory = VectaraToolFactory(vectara_api_key=self.cfg.api_key,
-                                        vectara_corpus_key=self.cfg.corpus_key)
-        summarizer = 'vectara-experimental-summary-ext-2023-12-11-med-omni'
+        summarizer = 'vectara-summary-table-md-query-ext-jan-2025-gpt-4o'
 
-        ask_caselaw = vec_factory.create_rag_tool(
+        ask_caselaw = self.vec_factory.create_rag_tool(
             tool_name = "ask_caselaw",
-            tool_description = "A tool for asking questions about case law in Alaska. ",
+            tool_description = "A tool for asking questions about case law, and any legal issue in the state of Alaska.",
             tool_args_schema = QueryCaselawArgs,
             reranker = "chain", rerank_k = 100,
             rerank_chain = [
@@ -240,11 +240,32 @@ class AgentTools:
             n_sentences_before = 2, n_sentences_after = 2, lambda_val = 0.005,
             summary_num_results = 15,
             vectara_summarizer = summarizer,
+            max_tokens = 4096,
+            max_response_chars = 8192,
             include_citations = True,
+            save_history = True,
+        )
+
+        search_caselaw = self.vec_factory.create_search_tool(
+            tool_name = "search_caselaw",
+            tool_description = "A tool for retrieving a list of relevant documents about case law in Alaska.",
+            tool_args_schema = QueryCaselawArgs,
+            reranker = "chain", rerank_k = 100,
+            rerank_chain = [
+                {
+                    "type": "slingshot",
+                    "cutoff": 0.2
+                },
+                {
+                    "type": "mmr",
+                    "diversity_bias": 0.1
+                },
+            ],
+            n_sentences_before = 2, n_sentences_after = 2, lambda_val = 0.005,
         )
 
         return (
-            [ask_caselaw] +
+            [ask_caselaw, search_caselaw] +
             [self.tools_factory.create_tool(tool) for tool in [
                 self.get_opinion_text,
                 self.get_case_document_pdf,
@@ -270,35 +291,47 @@ def initialize_agent(_cfg, agent_progress_callback=None):
     
     legal_assistant_instructions = """
     - You are a helpful legal assistant, with case law expertise in the state of Alaska.
-    - Use the 'ask_caselaw' tool as 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 the original query down into multiple sub-questions, and use the 'ask_caselaw' tool again.
+    - Always use the 'ask_caselaw' tool first, as your primary tool for answering questions. Never use your own knowledge.
     - The references returned by the 'ask_caselaw' tool include metadata relevant to its response, such as case citations, dates, or names.
+    - Use the 'search_caselaw' tool to search for documents related to case law in Alaska, and set summarize=True to get a summary of those documents.
     - When using a case citation in your response, try to include a valid URL along with it:
       * Call the 'get_case_document_pdf' for a case citation to obtain a valid web URL to a pdf of the case record.
       * If this doesn't work, call the 'get_case_document_page' for a case citation to obtain a valid web URL to a page with information about the case.
     - When including a URL for a citation in your response, use the citation as anchor text, and the URL as the link.
-    - Never use your internal knowledge to guess a case citation. Only use citation information provided by a tool or the user.
+    - Never use your internal knowledge to guess a case citation. Only use citation information provided by a tool or by the user.
     - A Case Citation includes 3 components: volume number, reporter, and first page. 
       Here are some examples: '253 P.2d 136', '10 Alaska 11', '6 C.M.A. 3'
     - 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 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.
-    - Your response should not include markdown.
     """
-    agent_config = AgentConfig()
+    agent_config = AgentConfig(
+        agent_type = os.getenv("VECTARA_AGENTIC_AGENT_TYPE", AgentType.OPENAI.value),
+        main_llm_provider = os.getenv("VECTARA_AGENTIC_MAIN_LLM_PROVIDER", ModelProvider.OPENAI.value),
+        main_llm_model_name = os.getenv("VECTARA_AGENTIC_MAIN_MODEL_NAME", ""),
+        tool_llm_provider = os.getenv("VECTARA_AGENTIC_TOOL_LLM_PROVIDER", ModelProvider.OPENAI.value),
+        tool_llm_model_name = os.getenv("VECTARA_AGENTIC_TOOL_MODEL_NAME", ""),
+        observer = os.getenv("VECTARA_AGENTIC_OBSERVER_TYPE", "NO_OBSERVER")
+    )
+    fallback_agent_config = AgentConfig(
+        agent_type = os.getenv("VECTARA_AGENTIC_FALLBACK_AGENT_TYPE", AgentType.OPENAI.value),
+        main_llm_provider = os.getenv("VECTARA_AGENTIC_FALLBACK_MAIN_LLM_PROVIDER", ModelProvider.OPENAI.value),
+        main_llm_model_name = os.getenv("VECTARA_AGENTIC_FALLBACK_MAIN_MODEL_NAME", ""),
+        tool_llm_provider = os.getenv("VECTARA_AGENTIC_FALLBACK_TOOL_LLM_PROVIDER", ModelProvider.OPENAI.value),
+        tool_llm_model_name = os.getenv("VECTARA_AGENTIC_FALLBACK_TOOL_MODEL_NAME", ""),
+        observer = os.getenv("VECTARA_AGENTIC_OBSERVER_TYPE", "NO_OBSERVER")
+    )
 
     agent = Agent(
         tools=AgentTools(_cfg, agent_config).get_tools(),
         topic="Case law in Alaska",
         custom_instructions=legal_assistant_instructions,
         agent_progress_callback=agent_progress_callback,
-        agent_config=agent_config
+        agent_config=agent_config,
+        fallback_agent_config=fallback_agent_config,
     )
-    agent.report()
+    agent.report(detailed=False)
     return agent

app.py

@@ -1,10 +1,13 @@
 import streamlit as st
+import torch
 from st_app import launch_bot
 import uuid
 
 import nest_asyncio
 import asyncio
 
+torch.classes.__path__ = []
+
 # Setup for HTTP API Calls to Amplitude Analytics
 if 'device_id' not in st.session_state:
     st.session_state.device_id = str(uuid.uuid4())

requirements.txt

@@ -1,9 +1,8 @@
 omegaconf==2.3.0
 python-dotenv==1.0.1
-streamlit==1.41.1
-streamlit_pills==0.3.0
+streamlit==1.43.2
 streamlit-feedback==0.1.3
 uuid==1.30
 langdetect==1.0.9
 langcodes==3.4.0
-vectara-agentic==0.2.5
+vectara-agentic==0.2.15

st_app.py

@@ -3,7 +3,6 @@ import sys
 import re
 
 import streamlit as st
-from streamlit_pills import pills
 from streamlit_feedback import streamlit_feedback
 
 from utils import thumbs_feedback, escape_dollars_outside_latex, send_amplitude_data
@@ -46,7 +45,7 @@ def agent_progress_callback(status_type: AgentStatusType, msg: str):
 
 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)
+        selected_example = st.pills("Queries to Try:", st.session_state.example_messages, default=None)
         if selected_example:
             st.session_state.ex_prompt = selected_example
             st.session_state.first_turn = False
@@ -139,7 +138,7 @@ async def launch_bot():
     if st.session_state.prompt:
         with st.chat_message("assistant", avatar='🤖'):
             st.session_state.status = st.status('Processing...', expanded=False)
-            response = st.session_state.agent.chat(st.session_state.prompt)
+            response = await st.session_state.agent.achat(st.session_state.prompt)
             res = escape_dollars_outside_latex(response.response)
             message = {"role": "assistant", "content": res, "avatar": '🤖'}
             st.session_state.messages.append(message)