app.py

from collections import defaultdict
import json
import os
import re

from langchain.tools.retriever import create_retriever_tool
from langchain.agents import AgentExecutor
from langchain.agents import create_openai_tools_agent
from langchain.agents.format_scratchpad.openai_tools import format_to_openai_tool_messages
from langchain.agents.output_parsers.openai_tools import OpenAIToolsAgentOutputParser
from langchain_core.documents import Document
from langchain_core.prompts import PromptTemplate
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.prompts import MessagesPlaceholder
from langchain_core.messages import AIMessage
from langchain_core.messages import HumanMessage
from langchain_core.runnables import RunnableParallel
from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser
from langchain_community.callbacks import get_openai_callback
from langchain_community.callbacks import StreamlitCallbackHandler
from langchain_community.embeddings import HuggingFaceBgeEmbeddings
from langchain_community.vectorstores.utils import DistanceStrategy
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic
from langchain_pinecone import PineconeVectorStore
from pinecone import Pinecone
import streamlit as st


st.set_page_config(layout="wide", page_title="LegisQA")

os.environ["LANGCHAIN_API_KEY"] = st.secrets["langchain_api_key"]
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_PROJECT"] = st.secrets["langchain_project"]
os.environ["TOKENIZERS_PARALLELISM"] = "false"


SS = st.session_state
SEED = 292764
CONGRESS_NUMBERS = [113, 114, 115, 116, 117, 118]
SPONSOR_PARTIES = ["D", "R", "L", "I"]
CONGRESS_GOV_TYPE_MAP = {
    "hconres": "house-concurrent-resolution",
    "hjres": "house-joint-resolution",
    "hr": "house-bill",
    "hres": "house-resolution",
    "s": "senate-bill",
    "sconres": "senate-concurrent-resolution",
    "sjres": "senate-joint-resolution",
    "sres": "senate-resolution",
}
OPENAI_CHAT_MODELS = [
    "gpt-3.5-turbo-0125",
    "gpt-4-0125-preview",
]
ANTHROPIC_CHAT_MODELS = [
    "claude-3-opus-20240229",
    "claude-3-sonnet-20240229",
    "claude-3-haiku-20240307",
]
CHAT_MODELS = OPENAI_CHAT_MODELS + ANTHROPIC_CHAT_MODELS


def get_sponsor_url(bioguide_id: str) -> str:
    return f"https://bioguide.congress.gov/search/bio/{bioguide_id}"


def get_congress_gov_url(congress_num: int, legis_type: str, legis_num: int) -> str:
    lt = CONGRESS_GOV_TYPE_MAP[legis_type]
    return f"https://www.congress.gov/bill/{int(congress_num)}th-congress/{lt}/{int(legis_num)}"


def load_bge_embeddings():
    model_name = "BAAI/bge-small-en-v1.5"
    model_kwargs = {"device": "cpu"}
    encode_kwargs = {"normalize_embeddings": True}
    emb_fn = HuggingFaceBgeEmbeddings(
        model_name=model_name,
        model_kwargs=model_kwargs,
        encode_kwargs=encode_kwargs,
        query_instruction="Represent this question for searching relevant passages: ",
    )
    return emb_fn


def load_pinecone_vectorstore():
    emb_fn = load_bge_embeddings()
    vectorstore = PineconeVectorStore(
        embedding=emb_fn,
        text_key="text",
        distance_strategy=DistanceStrategy.COSINE,
        pinecone_api_key=st.secrets["pinecone_api_key"],
        index_name=st.secrets["pinecone_index_name"],
    )
    return vectorstore


def render_outreach_links():
    nomic_base_url = "https://atlas.nomic.ai/data/gabrielhyperdemocracy"
    nomic_map_name = "us-congressional-legislation-s1024o256nomic"
    nomic_url = f"{nomic_base_url}/{nomic_map_name}/map"
    hf_url = "https://huggingface.co/hyperdemocracy"
    pc_url = "https://www.pinecone.io/blog/serverless"
    st.subheader(":brain: About [hyperdemocracy](https://hyperdemocracy.us)")
    st.subheader(f":world_map: Visualize [nomic atlas]({nomic_url})")
    st.subheader(f":hugging_face: Raw [huggingface datasets]({hf_url})")
    st.subheader(f":evergreen_tree: Index [pinecone serverless]({pc_url})")


def group_docs(docs) -> list[tuple[str, list[Document]]]:
    doc_grps = defaultdict(list)

    # create legis_id groups
    for doc in docs:
        doc_grps[doc.metadata["legis_id"]].append(doc)

    # sort docs in each group by start index
    for legis_id in doc_grps.keys():
        doc_grps[legis_id] = sorted(
            doc_grps[legis_id],
            key=lambda x: x.metadata["start_index"],
        )

    # sort groups by number of docs
    doc_grps = sorted(
        tuple(doc_grps.items()),
        key=lambda x: -len(x[1]),
    )

    return doc_grps


def format_docs(docs):
    """JSON grouped"""

    doc_grps = group_docs(docs)
    out = []
    for legis_id, doc_grp in doc_grps:
        dd = {
            "legis_id": doc_grp[0].metadata["legis_id"],
            "title": doc_grp[0].metadata["title"],
            "sponsor": doc_grp[0].metadata["sponsor_full_name"],
            "snippets": [doc.page_content for doc in doc_grp],
        }
        out.append(dd)
    return json.dumps(out, indent=4)


def escape_markdown(text):
    MD_SPECIAL_CHARS = r"\`*_{}[]()#+-.!$"
    for char in MD_SPECIAL_CHARS:
        text = text.replace(char, "\\" + char)
    return text


def get_vectorstore_filter():
    vs_filter = {}
    if SS["filter_legis_id"] != "":
        vs_filter["legis_id"] = SS["filter_legis_id"]
    if SS["filter_bioguide_id"] != "":
        vs_filter["sponsor_bioguide_id"] = SS["filter_bioguide_id"]
    vs_filter = {**vs_filter, "congress_num": {"$in": SS["filter_congress_nums"]}}
    vs_filter = {**vs_filter, "sponsor_party": {"$in": SS["filter_sponsor_parties"]}}
    return vs_filter


def render_doc_grp(legis_id: str, doc_grp: list[Document]):
    first_doc = doc_grp[0]

    congress_gov_url = get_congress_gov_url(
        first_doc.metadata["congress_num"],
        first_doc.metadata["legis_type"],
        first_doc.metadata["legis_num"],
    )
    congress_gov_link = f"[congress.gov]({congress_gov_url})"


    ref = "{} chunks from {}\n\n{}\n\n{}\n\n[{} ({}) ]({})".format(
        len(doc_grp),
        first_doc.metadata["legis_id"],
        first_doc.metadata["title"],
        congress_gov_link,
        first_doc.metadata["sponsor_full_name"],
        first_doc.metadata["sponsor_bioguide_id"],
        get_sponsor_url(first_doc.metadata["sponsor_bioguide_id"]),
    )
    doc_contents = [
        "[start_index={}] ".format(int(doc.metadata["start_index"])) + doc.page_content
        for doc in doc_grp
    ]
    with st.expander(ref):
        st.write(escape_markdown("\n\n...\n\n".join(doc_contents)))


def legis_id_to_link(legis_id: str) -> str:
    congress_num, legis_type, legis_num = legis_id.split("-")
    return get_congress_gov_url(congress_num, legis_type, legis_num)


def legis_id_match_to_link(matchobj):
    mstring = matchobj.string[matchobj.start() : matchobj.end()]
    url = legis_id_to_link(mstring)
    link = f"[{mstring}]({url})"
    return link


def replace_legis_ids_with_urls(text):
    pattern = "11[345678]-[a-z]+-\d{1,5}"
    rtext = re.sub(pattern, legis_id_match_to_link, text)
    return rtext


def render_guide():

    st.write(
        """
When you send a query to LegisQA, it will attempt to retrieve relevant content from the past six congresses ([113th-118th](https://en.wikipedia.org/wiki/List_of_United_States_Congresses)) covering 2013 to the present, pass it to a [large language model (LLM)](https://en.wikipedia.org/wiki/Large_language_model), and generate a response. This technique is known as Retrieval Augmented Generation (RAG). You can read [an academic paper](https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html) or [a high level summary](https://research.ibm.com/blog/retrieval-augmented-generation-RAG) to get more details. Once the response is generated, the retrieved content will be available for inspection with links to the bills and sponsors.


## Disclaimer

This is a research project. The RAG technique helps to ground the LLM response by providing context from a trusted source, but it does not guarantee a high quality response. We encourage you to play around, find questions that work and find questions that fail. There is a small monthly budget dedicated to the OpenAI endpoints. Once that is used up each month, queries will no longer work.


## Sidebar Config

Use the `Generative Config` to change LLM parameters.
Use the `Retrieval Config` to change the number of chunks retrieved from our congress corpus and to apply various filters to the content before it is retrieved (e.g. filter to a specific set of congresses). Use the `Prompt Config` to try out different document formatting and prompting strategies.

    """
    )


def render_example_queries():

    with st.expander("Example Queries"):
        st.write(
            """

```
What are the themes around artificial intelligence?
```

```
Write a well cited 3 paragraph essay on food insecurity.
```

```
Create a table summarizing major climate change ideas with columns legis_id, title, idea.
```

```
Write an action plan to keep social security solvent.
```

```
Suggest reforms that would benefit the Medicaid program.
```

        """
        )


def render_sidebar():

    with st.container(border=True):
        render_outreach_links()

    st.checkbox("escape markdown in answer", key="response_escape_markdown")
    st.checkbox("add legis urls in answer", value=True, key="response_add_legis_urls")

    with st.expander("Generative Config"):
        st.selectbox(label="model name", options=CHAT_MODELS, key="model_name")
        st.slider(
            "temperature", min_value=0.0, max_value=2.0, value=0.0, key="temperature"
        )
        st.slider(
            "max_output_tokens", min_value=512, max_value=1024, key="max_output_tokens"
        )
        st.slider("top_p", min_value=0.0, max_value=1.0, value=1.0, key="top_p")

    with st.expander("Retrieval Config"):
        st.slider(
            "Number of chunks to retrieve",
            min_value=1,
            max_value=32,
            value=8,
            key="n_ret_docs",
        )
        st.text_input("Bill ID (e.g. 118-s-2293)", key="filter_legis_id")
        st.text_input("Bioguide ID (e.g. R000595)", key="filter_bioguide_id")
        st.multiselect(
            "Congress Numbers",
            CONGRESS_NUMBERS,
            default=CONGRESS_NUMBERS,
            key="filter_congress_nums",
        )
        st.multiselect(
            "Sponsor Party",
            SPONSOR_PARTIES,
            default=SPONSOR_PARTIES,
            key="filter_sponsor_parties",
        )


def render_query_rag_tab():

    render_example_queries()

    QUERY_TEMPLATE = """Use the following excerpts from US congressional legislation to respond to the user's query. The excerpts are formatted as a JSON list. Each JSON object has "legis_id", "title", "sponsor", and "snippets" keys. If a snippet is useful in writing part of your response, then cite the "title", "legis_id", and "sponsor" in the response. If you don't know how to respond, just tell the user.

---

Congressional Legislation Excerpts:

{context}

---

Query: {query}"""

    prompt = ChatPromptTemplate.from_messages(
        [
            ("system", "You are an expert legislative analyst."),
            ("human", QUERY_TEMPLATE),
        ]
    )

    with st.form("query_form"):
        st.text_area("Enter query:", key="query")
        query_submitted = st.form_submit_button("Submit")

    if query_submitted:

        vs_filter = get_vectorstore_filter()
        retriever = vectorstore.as_retriever(
            search_kwargs={"k": SS["n_ret_docs"], "filter": vs_filter},
        )

        rag_chain = (
            RunnableParallel(
                {
                    "docs": retriever,  # list of docs
                    "query": RunnablePassthrough(),  # str
                }
            )
            .assign(context=(lambda x: format_docs(x["docs"])))
            .assign(output=prompt | llm | StrOutputParser())
        )

        if SS["model_name"] in OPENAI_CHAT_MODELS:
            with get_openai_callback() as cb:
                SS["out"] = rag_chain.invoke(SS["query"])
                SS["cb"] = cb
        else:
            SS["out"] = rag_chain.invoke(SS["query"])

    if "out" in SS:

        out_display = SS["out"]["output"]
        if SS["response_escape_markdown"]:
            out_display = escape_markdown(out_display)
        if SS["response_add_legis_urls"]:
            out_display = replace_legis_ids_with_urls(out_display)
        with st.container(border=True):
            st.write("Response")
            st.info(out_display)

        if "cb" in SS:
            with st.container(border=True):
                st.write("API Usage")
                st.warning(SS["cb"])

        with st.container(border=True):
            doc_grps = group_docs(SS["out"]["docs"])
            st.write(
                "Retrieved Chunks (note that you may need to 'right click' on links in the expanders to follow them)"
            )
            for legis_id, doc_grp in doc_grps:
                render_doc_grp(legis_id, doc_grp)

        with st.expander("Debug"):
            st.write(SS["out"])


def render_query_agent_tab():

    from custom_tools import get_retriever_tool

    from langchain_community.tools import WikipediaQueryRun
    from langchain_community.utilities import WikipediaAPIWrapper
    from langchain.agents import load_tools
    from langchain.agents import create_react_agent
    from langchain import hub

    if SS["model_name"] not in OPENAI_CHAT_MODELS:
        st.write("only supported with OpenAI for now")
        return

    vs_filter = get_vectorstore_filter()
    retriever = vectorstore.as_retriever(
        search_kwargs={"k": SS["n_ret_docs"], "filter": vs_filter},
    )
    legis_retrieval_tool = get_retriever_tool(
        retriever,
        "search_legislation",
        "Searches and returns excerpts from congressional legislation. Always call this tool first.",
        format_docs,
    )

    api_wrapper = WikipediaAPIWrapper(top_k_results=4, doc_content_chars_max=800)
    wiki_search_tool = WikipediaQueryRun(api_wrapper=api_wrapper)

    ddg_tool = load_tools(["ddg-search"])[0]

    avatars = {"human": "user", "ai": "assistant"}
    tools = [legis_retrieval_tool, wiki_search_tool, ddg_tool]
    llm_with_tools = llm.bind_tools(tools)

    agent_prompt = ChatPromptTemplate.from_messages(
        [
            ("system", "You are a helpful assistant."),
            ("human", "{input}"),
            MessagesPlaceholder(variable_name="agent_scratchpad"),
        ]
    )
    agent = (
        {
            "input": lambda x: x["input"],
            "agent_scratchpad": lambda x: format_to_openai_tool_messages(
                x["intermediate_steps"]
            ),
        }
        | agent_prompt
        | llm_with_tools
        | OpenAIToolsAgentOutputParser()
    )

    prompt = hub.pull("hwchase17/react")
    agent = create_react_agent(llm, tools, prompt)
    agent_executor = AgentExecutor(
        agent=agent,
        tools=tools,
        return_intermediate_steps=True,
        handle_parsing_errors=True,
        verbose=True,
    )

    if user_input := st.chat_input(key="single_query_agent_input"):
        st.chat_message("user").write(user_input)
        with st.chat_message("assistant"):
            st_callback = StreamlitCallbackHandler(st.container())
            response = agent_executor.invoke({"input": user_input}, {"callbacks": [st_callback]})
            st.write(response["output"])


def render_chat_agent_tab():
    st.write("Coming Soon")





##################


st.title(
    ":classical_building: LegisQA - Chat With Congressional Bills :classical_building:"
)


with st.sidebar:
    render_sidebar()


if SS["model_name"] in OPENAI_CHAT_MODELS:
    llm = ChatOpenAI(
        model_name=SS["model_name"],
        temperature=SS["temperature"],
        openai_api_key=st.secrets["openai_api_key"],
        model_kwargs={"top_p": SS["top_p"], "seed": SEED},
        max_tokens=SS["max_output_tokens"],
    )
elif SS["model_name"] in ANTHROPIC_CHAT_MODELS:
    llm = ChatAnthropic(
        model_name=SS["model_name"],
        temperature=SS["temperature"],
        anthropic_api_key=st.secrets["anthropic_api_key"],
        top_p=SS["top_p"],
        max_tokens_to_sample=SS["max_output_tokens"],
    )
else:
    raise ValueError()


vectorstore = load_pinecone_vectorstore()

query_rag_tab, query_agent_tab, chat_agent_tab, guide_tab = st.tabs([
    "query_rag",
    "query_agent",
    "chat_agent",
    "guide",
])

with query_rag_tab:
    render_query_rag_tab()

with query_agent_tab:
    render_query_agent_tab()

with chat_agent_tab:
    render_chat_agent_tab()

with guide_tab:
    render_guide()