hf-legisqa / app.py
gabrielaltay's picture
more text
6bfad85
raw
history blame
12.7 kB
from collections import defaultdict
import json
from langchain_core.documents import Document
from langchain_core.prompts import PromptTemplate
from langchain_core.runnables import RunnableParallel
from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser
from langchain_community.embeddings import HuggingFaceBgeEmbeddings
from langchain_community.vectorstores.utils import DistanceStrategy
from langchain_openai import ChatOpenAI
from langchain_pinecone import PineconeVectorStore
from pinecone import Pinecone
import streamlit as st
st.set_page_config(layout="wide", page_title="LegisQA")
SS = st.session_state
SEED = 292764
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",
]
PREAMBLE = "You are an expert analyst. Use the following excerpts from US congressional legislation to respond to the user's query."
PROMPT_TEMPLATES = {
"v1": PREAMBLE
+ """ If you don't know how to respond, just tell the user.
{context}
Question: {question}""",
"v2": PREAMBLE
+ """ Each snippet starts with a header that includes a unique snippet number (snippet_num), a legis_id, and a title. Your response should reference particular snippets using legis_id and title. If you don't know how to respond, just tell the user.
{context}
Question: {question}""",
"v3": PREAMBLE
+ """ Each excerpt starts with a header that includes a legis_id, and a title followed by one or more text snippets. When using text snippets in your response, you should mention the legis_id and title. If you don't know how to respond, just tell the user.
{context}
Question: {question}""",
"v4": PREAMBLE
+ """ The excerpts are formatted as a JSON list. Each JSON object has "legis_id", "title", and "snippets" keys. If a snippet is useful in writing part of your response, then mention the "title" and "legis_id" inline as you write. If you don't know how to respond, just tell the user.
{context}
Query: {question}""",
}
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 get_govtrack_url(congress_num: int, legis_type: str, legis_num: int) -> str:
return f"https://www.govtrack.us/congress/bills/{int(congress_num)}/{legis_type}{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()
pc = Pinecone(api_key=st.secrets["pinecone_api_key"])
index = pc.Index(st.secrets["pinecone_index_name"])
vectorstore = PineconeVectorStore(
index=index,
embedding=emb_fn,
text_key="text",
distance_strategy=DistanceStrategy.COSINE,
)
return vectorstore
def write_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_v1(docs):
"""Simple double new line join"""
return "\n\n".join([doc.page_content for doc in docs])
def format_docs_v2(docs):
"""Format with snippet_num, legis_id, and title"""
def format_doc(idoc, doc):
return "snippet_num: {}\nlegis_id: {}\ntitle: {}\n... {} ...\n".format(
idoc,
doc.metadata["legis_id"],
doc.metadata["title"],
doc.page_content,
)
snips = []
for idoc, doc in enumerate(docs):
txt = format_doc(idoc, doc)
snips.append(txt)
return "\n===\n".join(snips)
def format_docs_v3(docs):
def format_header(doc):
return "legis_id: {}\ntitle: {}".format(
doc.metadata["legis_id"],
doc.metadata["title"],
)
def format_content(doc):
return "... {} ...\n".format(
doc.page_content,
)
snips = []
doc_grps = group_docs(docs)
for legis_id, doc_grp in doc_grps:
first_doc = doc_grp[0]
head = format_header(first_doc)
contents = []
for idoc, doc in enumerate(doc_grp):
txt = format_content(doc)
contents.append(txt)
snips.append("{}\n\n{}".format(head, "\n".join(contents)))
return "\n===\n".join(snips)
def format_docs_v4(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"],
"snippets": [doc.page_content for doc in doc_grp],
}
out.append(dd)
return json.dumps(out, indent=4)
DOC_FORMATTERS = {
"v1": format_docs_v1,
"v2": format_docs_v2,
"v3": format_docs_v3,
"v4": format_docs_v4,
}
def escape_markdown(text):
MD_SPECIAL_CHARS = r"\`*_{}[]()#+-.!$"
for char in MD_SPECIAL_CHARS:
text = text.replace(char, "\\" + char)
return text
st.title(":classical_building: LegisQA :classical_building:")
st.header("Explore Congressional Legislation")
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.
This 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. Try different models. Find questions that work and find questions that fail.""")
st.header("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 the major climate change ideas with columns legis_id, title, idea.
```
"""
)
with st.sidebar:
with st.container(border=True):
write_outreach_links()
st.checkbox("escape markdown in answer", key="response_escape_markdown")
with st.expander("Generative Config"):
st.selectbox(label="model name", options=OPENAI_CHAT_MODELS, key="model_name")
st.slider(
"temperature", min_value=0.0, max_value=2.0, value=0.0, key="temperature"
)
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=40,
value=10,
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.text_input("Congress (e.g. 118)", key="filter_congress_num")
with st.expander("Prompt Config"):
st.selectbox(
label="prompt version",
options=PROMPT_TEMPLATES.keys(),
index=3,
key="prompt_version",
)
st.text_area(
"prompt template",
PROMPT_TEMPLATES[SS["prompt_version"]],
height=300,
key="prompt_template",
)
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},
)
vectorstore = load_pinecone_vectorstore()
format_docs = DOC_FORMATTERS[SS["prompt_version"]]
with st.form("my_form"):
st.text_area("Enter query:", key="query")
query_submitted = st.form_submit_button("Submit")
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"]
if SS["filter_congress_num"] != "":
vs_filter["congress_num"] = int(SS["filter_congress_num"])
return vs_filter
if query_submitted:
vs_filter = get_vectorstore_filter()
retriever = vectorstore.as_retriever(
search_kwargs={"k": SS["n_ret_docs"], "filter": vs_filter},
)
prompt = PromptTemplate.from_template(SS["prompt_template"])
rag_chain_from_docs = (
RunnablePassthrough.assign(context=(lambda x: format_docs(x["context"])))
| prompt
| llm
| StrOutputParser()
)
rag_chain_with_source = RunnableParallel(
{"context": retriever, "question": RunnablePassthrough()}
).assign(answer=rag_chain_from_docs)
out = rag_chain_with_source.invoke(SS["query"])
SS["out"] = out
def write_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})"
gov_track_url = get_govtrack_url(
first_doc.metadata["congress_num"],
first_doc.metadata["legis_type"],
first_doc.metadata["legis_num"],
)
gov_track_link = f"[govtrack.us]({gov_track_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,
gov_track_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)))
out = SS.get("out")
if out:
if SS["response_escape_markdown"]:
st.info(escape_markdown(out["answer"]))
else:
st.info(out["answer"])
doc_grps = group_docs(out["context"])
for legis_id, doc_grp in doc_grps:
write_doc_grp(legis_id, doc_grp)
with st.expander("Debug doc format"):
st.text_area("formatted docs", value=format_docs(out["context"]), height=600)
# st.write(json.loads(format_docs(out["context"])))