from haystack import Document from haystack.document_stores import InMemoryDocumentStore, ElasticsearchDocumentStore, FAISSDocumentStore from haystack.nodes import BM25Retriever from haystack.pipelines import DocumentSearchPipeline import pandas as pd import panel as pn import param pn.extension('tabulator') pn.extension(sizing_mode="scale_both") import hvplot.pandas # load data infile = "Agency Inventory AI Usage - Sheet1.tsv" df = pd.read_csv(infile, sep="\t", lineterminator='\n') # rearrange column order col_list = ['Agency', 'Name of Inventory Item', 'Primary Type of AI', 'Purpose of AI', 'Length of Usage', 'Does it directly impact the public?', 'Vendor System', 'Description of Inventory Item', 'Other Notes\r'] df = df[col_list] # remove trailing \r from 'Other Notes' header df = df.rename(columns = {'Other Notes\r' : 'Other Notes'}) # remove trailing spaces from agency names (caused duplicate instance of "DOC") df['Agency'] = df['Agency'].apply(lambda x : x.rstrip()) # columns not useful for filtering no_filter_cols = ['Name of Inventory Item', 'Description of Inventory Item', "Other Notes"] # columns to be used for filtering filter_cols = [c for c in df.columns.unique() if c not in no_filter_cols] # column selector for main plot plot_column_selector = pn.widgets.Select(options=filter_cols, name="Plot category: ") # agency selector for main plot plot_agency_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df["Agency"].unique()), value=["Select all"], name="Optional - filter by agency") # selectors below are all for interactive dataframe agency_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df["Agency"].unique()), value=["Select all"], name="Agency") type_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df['Primary Type of AI'].unique()), value=["Select all"], name='Primary Type of AI') purpose_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df["Purpose of AI"].unique()), value=["Select all"], name="Purpose of AI") length_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df['Length of Usage'].unique()), value=["Select all"], name="Length of Usage") impact_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df['Does it directly impact the public?'].unique()), value=["Select all"], name='Does it directly impact the public?') vendor_selector = pn.widgets.MultiSelect(options=["Select all"] + list(df['Vendor System'].unique()), value=["Select all"], name='Vendor System') row_filters = [agency_selector, type_selector, purpose_selector, length_selector, impact_selector, vendor_selector] def custom_plot(table, column_selector, agency_selector): if "Select all" not in agency_selector: table = table[table['Agency'].isin(agency_selector)] table = table[column_selector].value_counts().sort_values(ascending=True) return table.hvplot.barh(width=600, height=400, color="#336BCC") def custom_table_filter(table, agency_selector, type_selector, purpose_selector, length_selector, impact_selector, vendor_selector): """ This repetitive approach was the only way I could get things working with a 'Select all' menu option. """ if "Select all" not in agency_selector: table = table[table["Agency"].isin(agency_selector)] if "Select all" not in type_selector: table = table[table['Primary Type of AI'].isin(type_selector)] if "Select all" not in purpose_selector: table = table[table["Purpose of AI"].isin(purpose_selector)] if "Select all" not in length_selector: table = table[table['Length of Usage'].isin(length_selector)] if "Select all" not in impact_selector: table = table[table['Does it directly impact the public?'].isin(impact_selector)] if "Select all" not in vendor_selector: table = table[table['Vendor System'].isin(vendor_selector)] return table custom_table = pn.widgets.Tabulator(df, pagination="local", page_size=350, layout="fit_data", width=800, height=550) custom_table.add_filter(pn.bind(custom_table_filter, agency_selector=agency_selector, type_selector=type_selector, purpose_selector=purpose_selector, length_selector=length_selector, impact_selector=impact_selector, vendor_selector=vendor_selector)) interactive_plot = pn.bind(custom_plot, table=df, column_selector=plot_column_selector, agency_selector=plot_agency_selector) overview_stacked = pn.Column( pn.pane.Markdown(""" Plot shows the total count of entries, aggregated by various categories. Change the category with the dropdown menu. The total number of records in the database is 337, but some fields have missing values. In particular, 'Vendor System' and 'Primary Type of AI' were not always filled out."""), pn.Column(pn.Row(plot_column_selector, plot_agency_selector), pn.Row(interactive_plot, width=500)) ) overview_card = pn.Card(overview_stacked, header="# Overview of the data") filename, button = custom_table.download_menu( text_kwargs={'name': 'Enter filename ending in .csv or .json', 'value': 'default.csv'}, button_kwargs={'name': 'Download table'} ) download_card = pn.Card(pn.pane.Markdown(""" Download current table in .csv or .json format. File format will be automatically selected based on the file extension. """), filename, button, header="### Download") table_card = pn.Card( pn.Row( pn.Column( pn.pane.Markdown(""" ### Filter with the menus below """),pn.WidgetBox(*row_filters), styles=dict(background='#DDE6FF') ), pn.Column(pn.pane.Markdown(""" ### Scroll horizontally and vertically to see all data """), custom_table)), download_card, header="# Explore the data" ) # stacked bar plot of impact by agency (static plot) impact_counts = df.groupby('Agency')['Does it directly impact the public?'].value_counts() impact_counts = impact_counts.sort_index(level="Agency", ascending=False) impact_count_df = pd.DataFrame(impact_counts).rename(columns={'Does it directly impact the public?' : "Count"}) impact_plot = impact_count_df.hvplot.barh(stacked=True, width=500, height=400, color=[ "#019C6D", "#336BCC", "#F41903",], legend="bottom_right") impact_card = pn.Card( pn.Column( pn.pane.Markdown(""" Number of systems with no, indirect, or direct impact on the public. These judgements were made by Anna Blue and are unique to her report."""), impact_plot), header="# Impact on the public, by agency") # keyword search class TableIndices(param.Parameterized): row_indices = param.List() col_indices = param.List() def __call__(self): return (self.row_indices, self.col_indices) def run_search(text, pipeline): if text == "": return None res = pipeline.run(query=text, params={"Retriever": {"top_k": 10}}) relevant_results = [r for r in res['documents'] if r.score > 0.5] result_rows = [doc.meta['index'] for doc in relevant_results] result_cols = [doc.meta['column_header'] for doc in relevant_results] table_indices = TableIndices(row_indices=result_rows, col_indices=result_cols) return table_indices def produce_table(df, table_indices): if not table_indices: return None result_df = df.iloc[table_indices.row_indices, :] result_df = result_df.drop_duplicates() color_df = result_df.copy() color_df.loc[:,:] = '' for row, col in zip(table_indices.row_indices, table_indices.col_indices): color_df.loc[row, col] = 'background-color: yellow' result_tab = pn.widgets.Tabulator(result_df,pagination="local", page_size=350, layout="fit_data", width=800, height=300) # cell coloration is working, but does not update properly unless empty search is run in between; # otherwise it re-uses the most recent color scheme; maybe related to https://github.com/holoviz/panel/issues/3363 # result_tab.style.apply(lambda x: color_df, axis=None) # giving up for now return result_tab def make_search_pane(result_tab): if not result_tab: return None filename_2, button_2 = result_tab.download_menu( text_kwargs={'name': 'Enter filename ending in .csv or .json', 'value': 'default.csv'}, button_kwargs={'name': 'Download search results'}) search_download_card = pn.Card(pn.pane.Markdown(""" Download search results in .csv or .json format. File format will be automatically selected based on the file extension."""), filename_2, button_2, header="### Download") search_result = pn.Column(pn.pane.Markdown(""" ### Scroll horizontally and vertically (if needed) to see everything. """), result_tab, search_download_card) return search_result # which columns to search col_list = ['Name of Inventory Item', 'Primary Type of AI', 'Purpose of AI', 'Description of Inventory Item', 'Other Notes'] # create document store, where each string from any of the relevant columns is a doc # save the row index as metadata docs = [] indices = list(df.index.values) for col in col_list: values = df[col].tolist() assert len(indices) == len(values) for i, val in zip(indices, values): dictionary = {'content' : val, 'meta' : {"index": i, "column_header" : col} } docs.append(Document.from_dict(dictionary)) document_store = InMemoryDocumentStore(use_bm25=True) document_store.write_documents(docs) retriever = BM25Retriever(document_store=document_store) pipeline = DocumentSearchPipeline(retriever) text_input = pn.widgets.TextInput(name='Search', placeholder='Enter text here...') result_indices = pn.bind(run_search, text=text_input, pipeline=pipeline) result_table = pn.bind(produce_table, df=df, table_indices=result_indices) result_pane = pn.bind(make_search_pane, result_tab=result_table) search_card = pn.Card( pn.Column( pn.Row( text_input, pn.pane.Markdown(""" This will search text in the following columns: * Name of Inventory Item * Primary Type of AI * Purpose of AI * Description of Inventory Item * Other Notes This is a keyword search based on the BM25 algorithm as implemented in the Haystack python library. """)), pn.Row(result_pane), ), header="# Search the text" ) main_text = """ The data visualized here come from a report by Anna Blue, a Social Impact Fellow at the Responsible AI Institute. The report was released in May 2023. Some agencies have released updated inventories since then, which are not reflected here. Anna's report consolidated data released by individual government agencies in compliance with Executive Order 13960, which requires federal agencies to produce an annual inventory of their AI usage. See her [blog post](https://www.responsible.ai/post/federal-government-ai-use-cases) for additional details, including links to the original data sources. """ template = pn.template.FastListTemplate( title='U.S. Government Use of AI', main=[pn.pane.Markdown(main_text), pn.Row(overview_card,impact_card), pn.Row(table_card), pn.Row(search_card)], accent_base_color="#FFDAC2", header_background="#0037A2") template.servable()