File size: 10,433 Bytes
ca49a1b
 
674face
ca49a1b
 
 
 
 
 
 
 
 
 
e64f989
ca49a1b
e64f989
 
 
ca49a1b
e64f989
ca49a1b
 
 
 
 
bfa3ec5
 
1e6395c
 
 
 
29aa949
ca49a1b
 
e64f989
 
 
ca49a1b
91a2e1f
ca49a1b
 
 
 
 
 
 
 
 
f5548b0
 
 
 
3bb944e
76c2d0a
08b14ba
 
f5548b0
ca49a1b
1e6395c
ca49a1b
 
91a2e1f
ca49a1b
 
 
f5548b0
1e6395c
f5548b0
ca49a1b
 
1e6395c
 
 
 
 
29aa949
1e6395c
f5548b0
1e6395c
 
ca49a1b
 
e01cf86
6193768
 
 
 
47a95ac
2cfbdba
6193768
 
8e99f61
6193768
a7844e4
2cfbdba
8e99f61
a7844e4
 
e5f6885
 
 
29aa949
e5f6885
2cfbdba
9f8d49e
e64f989
 
 
 
 
 
 
 
 
 
 
 
 
 
 
184544f
b620d1c
e64f989
 
 
 
 
 
 
 
ca49a1b
a7844e4
6193768
d7e50c8
8711c74
 
ca49a1b
 
 
5441dfc
 
ca49a1b
 
 
 
 
 
8f7a4f3
 
7025a94
ab933e6
 
 
 
 
 
8f7a4f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ca49a1b
e64f989
82ef5f2
 
e64f989
 
 
 
 
 
 
 
 
 
 
82ef5f2
e64f989
 
 
 
ca49a1b
 
8f7a4f3
 
ca49a1b
 
 
 
 
 
 
 
e64f989
9918933
8f7a4f3
 
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
import streamlit as st
import os
import re
import json
from dotenv import load_dotenv
from haystack.nodes.prompt import PromptNode, PromptTemplate
from haystack.nodes import EmbeddingRetriever
from haystack import Pipeline
import numpy as np
import pandas as pd
from haystack.document_stores import FAISSDocumentStore
from haystack.nodes import EmbeddingRetriever
from haystack.schema import Document
from huggingface_hub import login, HfApi, hf_hub_download, InferenceClient

# Get HF token
hf_token = os.environ["HF_TOKEN"]
login(token=hf_token, add_to_git_credential=True)

# Get openai API key
openai_key = os.environ["OPENAI_API_KEY"]


template = PromptTemplate(
    prompt="""
Answer the given question using the following documents. \
Formulate your answer in the style of an academic report. \
Provide example quotes and citations using extracted text from the documents. \
Use facts and numbers from the documents in your answer. \
Reference information used from documents at the end of each applicable sentence (ex: [source: document_name]), where 'document_name' is the text provided at the start of each document (demarcated by '- &&&' and '&&&:')'. \
If no relevant information to answer the question is present in the documents, just say you don't have enough information to answer. \
Context: {' - '.join(['&&& '+d.meta['document']+' ref. '+str(d.meta['ref_id'])+' &&&: '+d.content for d in documents])}; Question: {query}; Answer:""",
)

# Create a list of options for the dropdown
model_options = ['chatGPT','Llama2']

# Create a list of options for the dropdown
country_options = ['All Countries','Angola','Botswana','Lesotho','Kenya','Malawi','Mozambique','Namibia','Rwanda','South Africa','Zambia','Zimbabwe']

# List of examples
examples = [
    "-",
    "What specific initiatives are presented in the context to address the needs of groups such women and children to the effects climate change?",
    "In addition to gender, children, and youth, is there any mention of other groups facing disproportional impacts from climate change due to their geographic location, socio-economic status, age, gender, health, and occupation?"
]

def get_docs(input_query, country = None):
    '''
    Construct a hacky query to focus the retriever on the target country (see notes below)
    We take the top 150 k because we want to make sure we have 10 pertaining to the selected country
    '''
    if country == 'All Countries':
        query = input_query
    else:
        query = "For the country of "+country+", "+input_query
    # Retrieve top k documents
    docs = retriever.retrieve(query=query,top_k = 150)
    # Break out the key fields and convert to pandas for filtering
    docs = [{**x.meta,"score":x.score,"content":x.content} for x in docs]
    df_docs = pd.DataFrame(docs)
    if country != 'All Countries':
      df_docs = df_docs.query('country in @country')
    # Take the top 10
    df_docs = df_docs.head(10)
    # Get ourselves an index setup from which to base the source reference number from (in the prompt and matching afterwards)
    df_docs = df_docs.reset_index()
    df_docs['ref_id'] = df_docs.index + 1 # start the index at 1
    # Convert back to Document format
    ls_dict = []
    # Iterate over df and add relevant fields to the dict object
    for index, row in df_docs.iterrows():
        # Create a Document object for each row
        doc = Document(
            row['content'],
            meta={'country': row['country'],'document': row['document'], 'page': row['page'], 'file_name': row['file_name'], 'ref_id': row['ref_id'], 'score': row['score']}
        )

        # Append the Document object to the documents list
        ls_dict.append(doc)
    return(ls_dict)

def get_refs(docs, res):
  '''
  Parse response for engineered reference ids (refer to prompt template)
  Extract documents using reference ids
  '''
  res = res.lower()  # Convert to lowercase for matching
  # This pattern should be returned by gpt3.5
  # pattern = r'ref\. (\d+)\]\.'
  pattern = r'ref\. (\d+)'
  ref_ids = [int(match) for match in re.findall(pattern, res)]
  # extract
  result_str = ""  # Initialize an empty string to store the result
  for i in range(len(docs)):
      doc = docs[i].to_dict()
      ref_id = doc['meta']['ref_id']
      if ref_id in ref_ids:
        if doc['meta']['document'] == "Supplementary":
            result_str += "**Ref. " + str(ref_id) + " [" + doc['meta']['country'] + " " + doc['meta']['document'] + ':' + doc['meta']['file_name'] + ' p' + str(doc['meta']['page']) + "]:** " + "*'" + doc['content'] + "'*<br> <br>" # Add <br> for a line break
        else:
            result_str += "**Ref. " + str(ref_id) + " [" + doc['meta']['country'] + " " + doc['meta']['document'] + ' p' + str(doc['meta']['page']) + "]:** " + "*'" + doc['content'] + "'*<br> <br>" # Add <br> for a line break
  
  return result_str

# define a special function for putting the prompt together (as we can't use haystack)
def get_prompt_llama2(docs, query):
  base_prompt="Answer the given question using the following documents. \
  Formulate your answer in the style of an academic report. \
  Provide example quotes and citations using extracted text from the documents. \
  Use facts and numbers from the documents in your answer. \
  ALWAYS include references for information used from documents at the end of each applicable sentence using the format: '[ref. #]', where '[ref. #]' is included in the text provided at the start of each document (demarcated by the pattern '- &&& [ref. #] document_name &&&:')'. \
  Do not include page numbers in the references. \
  If no relevant information to answer the question is present in the documents, just say you don't have enough information to answer."
  # Add the meta data for references
  context = ' - '.join(['&&& [ref. '+str(d.meta['ref_id'])+'] '+d.meta['document_name']+' &&&: '+d.content for d in docs])
  prompt = base_prompt+"; Context: "+context+"; Question: "+query+"; Answer:"
  return(prompt)

def run_query(input_text, country, model_sel):
    docs = get_docs(input_text, country)
    # st.write('Selected country: ', country) # Debugging country
    if model_sel == "chatGPT":
        res = pipe.run(query=input_text, documents=docs)
        output = res["results"][0]
        references = get_refs(docs, res["results"][0])
    else:
        res = client.text_generation(get_prompt_llama2(docs, query=input_text), max_new_tokens=4000, temperature=0.01, model=model)
        output = res
        references = get_refs(docs, res)
    st.write('Response')
    st.success(output)
    st.write('References')
    st.markdown('References are based on text automatically extracted from climate policy documents. These extracts may contain non-legible characters or disjointed text as an artifact of the extraction procedure')
    st.markdown(references, unsafe_allow_html=True)

# Setup retriever, pulling from local faiss datastore
retriever = EmbeddingRetriever(
    document_store=FAISSDocumentStore.load(
        index_path="./cpv_full_southern_africa_kenya.faiss",
        config_path="./cpv_full_southern_africa_kenya.json",
    ),
    embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
    model_format="sentence_transformers",
    progress_bar=False,
)

with st.sidebar:
    # Dropdown selectbox
    country = st.sidebar.selectbox('Select a country:', country_options)
st.markdown(
    """
    * *For a comparative analysis of multiple countries, select 'All Countries'*
    * *Then be sure to mention the country names of interest in your query.*
    """
)
    # choice = st.sidebar.radio(label = 'Select the Document',
    #                         help = 'You can upload the document \
    #                         or else you can try a example document', 
    #                         options = ('Upload Document', 'Try Example'), 
    #                         horizontal = True)

with st.container():
        st.markdown("<h2 style='text-align: center;'> Climate Policy Documents: Vulnerabilities Analysis Q&A </h2>", unsafe_allow_html=True)
        st.write(' ')

with st.expander("ℹ️ - About this app", expanded=False):
    st.write(
        """
        This tool seeks to provide an interface for quering national climate policy documents (NDCs, LTS etc.). The current version is powered by chatGPT (3.5). The document store is limited to 10 Southern African countries (Angola, Botswana, Eswatini, Lesotho, Malawi, Mozambique, Namibia, South Africa, Zambia, Zimbabwe), as well as Kenya and Rwanda. The intended use case is to allow users to interact with the documents and obtain valuable insights on various vulnerable groups affected by climate change.
        
        **DISCLAIMER:** *This prototype tool based on LLMs (Language Models) is provided "as is" for experimental and exploratory purposes only, and should not be used for critical or production applications. Users are advised that the tool may contain errors, bugs, or limitations and should be used with caution and awareness of potential risks, and the developers make no warranties or guarantees regarding its performance, reliability, or suitability for any specific purpose.*
        """)

    # Display the text passages as radio buttons
    selected_example = st.radio("Example questions", examples)




# Dropdown selectbox: model
# model_sel = st.selectbox('Select an LLM:', model_options)
model_sel = "chatGPT"

#----Model Select logic-------
if model_sel == "chatGPT":
    model_name = "gpt-3.5-turbo"
    # Initialize the PromptNode
    pn = PromptNode(model_name_or_path=model_name, default_prompt_template=template, api_key=openai_key, max_length=2000, model_kwargs={"generation_kwargs": {"do_sample": False, "temperature": 0}})
    
    # Initialize the pipeline
    pipe = Pipeline()
    pipe.add_node(component=pn, name="prompt_node", inputs=["Query"])
else:
    # Currently disabled
    model = "meta-llama/Llama-2-70b-chat-hf"
    # Instantiate the inference client
    client = InferenceClient()






if selected_example == "-":
    text = st.text_area('Enter your question in the text box below using natural language or select an example from above:')
else:
    text = st.text_area('Enter your question in the text box below using natural language or select an example from above:', value=selected_example)


if st.button('Submit'):
    run_query(text, country=country, model_sel=model_sel)