File size: 12,670 Bytes
80275c5
 
 
 
 
 
 
 
249fca7
85952ea
d967c00
80275c5
 
a3e054f
 
 
80275c5
 
23a84f2
80275c5
23a84f2
 
 
 
 
 
 
 
 
 
 
 
 
 
d967c00
23a84f2
 
80275c5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1d2c336
23a84f2
 
 
 
 
 
 
 
 
1d2c336
 
 
 
 
80275c5
85952ea
80275c5
85952ea
23a84f2
85952ea
 
23a84f2
 
 
 
80275c5
 
 
 
 
6bfad85
 
 
 
 
80275c5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6bfad85
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80275c5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23a84f2
d3e2171
80275c5
 
 
6bfad85
80275c5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
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"])))