import json import os import gradio as gr import requests from huggingface_hub import HfApi hf_api = HfApi() roots_datasets = { dset.id.split("/")[-1]: dset for dset in hf_api.list_datasets( author="bigscience-data", use_auth_token=os.environ.get("bigscience_data_token") ) } def get_docid_html(docid): data_org, dataset, docid = docid.split("/") metadata = roots_datasets[dataset] if metadata.private: docid_html = ( f"🔒{dataset}/{docid}' ) else: docid_html = ( f"{dataset}/{docid}' ) return docid_html PII_TAGS = {"KEY", "EMAIL", "USER", "IP_ADDRESS", "ID", "IPv4", "IPv6"} PII_PREFIX = "PI:" def process_pii(text): for tag in PII_TAGS: text = text.replace( PII_PREFIX + tag, """REDACTED {}""".format( tag ), ) return text def process_results(results, highlight_terms): if len(results) == 0: return """

No results retrieved.



""" results_html = "" for result in results: tokens = result["text"].split() tokens_html = [] for token in tokens: if token in highlight_terms: tokens_html.append("{}".format(token)) else: tokens_html.append(token) tokens_html = " ".join(tokens_html) tokens_html = process_pii(tokens_html) meta_html = ( """

{}

""".format( result["meta"]["url"], result["meta"]["url"] ) if "meta" in result and result["meta"] is not None and "url" in result["meta"] else "" ) docid_html = get_docid_html(result["docid"]) results_html += """{}

Document ID: {}

Language: {}

{}


""".format( meta_html, docid_html, result["lang"], tokens_html ) return results_html + "
" def scisearch(query, language, num_results=10, exact_search=False): try: query = " ".join(query.split()) if query == "" or query is None: return "" post_data = {"query": query, "k": num_results, "exact_search": exact_search} if language != "detect_language": post_data["lang"] = language output = requests.post( os.environ.get("address"), headers={"Content-type": "application/json"}, data=json.dumps(post_data), timeout=60, ) payload = json.loads(output.text) if "err" in payload: if payload["err"]["type"] == "unsupported_lang": detected_lang = payload["err"]["meta"]["detected_lang"] return f"""

Detected language {detected_lang} is not supported.
Please choose a language from the dropdown or type another query.




""" results = payload["results"] highlight_terms = payload["highlight_terms"] if language == "detect_language": return ( ( f"""

Detected language: {results[0]["lang"]}




""" if len(results) > 0 and language == "detect_language" else "" ) + process_results(results, highlight_terms) ) if language == "all": results_html = "" for lang, results_for_lang in results.items(): if len(results_for_lang) == 0: results_html += f"""

No results for language: {lang}


""" continue collapsible_results = f"""
Results for language: {lang}
{process_results(results_for_lang, highlight_terms)}
""" results_html += collapsible_results return results_html return process_results(results, highlight_terms) except Exception as e: results_html = f"""

Raised {type(e).__name__}

Check if a relevant discussion already exists in the Community tab. If not, please open a discussion.

""" return results_html def perform_exact_search(query, num_results=10): try: print("perform_exact_search") query = " ".join(query.split()) if query == "" or query is None: return "" post_data = {"query": query, "k": num_results, "exact_search": True} print("post_data", post_data) output = requests.post( "http://34.105.160.81:8080", headers={"Content-type": "application/json"}, data=json.dumps(post_data), timeout=60, ) payload = json.loads(output.text) results = payload["results"] results_html = "" for result in results: print(result) result_html = """


""" query_start = result.find(query) query_end = query_start + len(query) result_html += result[0:query_start] result_html += "{}".format(result[query_start:query_end]) result_html += result[query_end:] results_html += result_html return results_html + "
" except Exception as e: results_html = f"""

Raised {type(e).__name__}

Check if a relevant discussion already exists in the Community tab. If not, please open a discussion.

""" def flag(query, language, num_results, issue_description): try: post_data = { "query": query, "k": num_results, "flag": True, "description": issue_description, } if language != "detect_language": post_data["lang"] = language output = requests.post( os.environ.get("address"), headers={"Content-type": "application/json"}, data=json.dumps(post_data), timeout=120, ) results = json.loads(output.text) except: print("Error flagging") return "" description = """#

🌸 🔎 ROOTS search tool 🔍 🌸

The ROOTS corpus was developed during the [BigScience workshop](https://bigscience.huggingface.co/) for the purpose of training the Multilingual Large Language Model [BLOOM](https://huggingface.co/bigscience/bloom). This tool allows you to search through the ROOTS corpus. We serve a BM25 index for each language or group of languages included in ROOTS. You can read more about the details of the tool design [here](https://huggingface.co/spaces/bigscience-data/scisearch/blob/main/roots_search_tool_specs.pdf). For more information and instructions on how to access the full corpus check [this form](https://forms.gle/qyYswbEL5kA23Wu99).""" if __name__ == "__main__": demo = gr.Blocks( css=".underline-on-hover:hover { text-decoration: underline; } .flagging { font-size:12px; color:Silver; }" ) with demo: with gr.Row(): gr.Markdown(value=description) with gr.Row(): query = gr.Textbox( lines=1, max_lines=1, placeholder="Type your query here...", label="Query", ) with gr.Row(): lang = gr.Dropdown( choices=[ "ar", "ca", "code", "en", "es", "eu", "fr", "id", "indic", "nigercongo", "pt", "vi", "zh", "detect_language", "all", ], value="en", label="Language", ) with gr.Row(): k = gr.Slider(1, 100, value=10, step=1, label="Max Results") with gr.Row(): with gr.Column(scale=1): exact_search = gr.Checkbox( value=False, label="Exact Search", variant="compact" ) with gr.Column(scale=4): submit_btn = gr.Button("Submit") with gr.Row(): results = gr.HTML(label="Results") flag_description = """

If you choose to flag your search, we will save the query, language and the number of results you requested. Please consider adding any additional context in the box on the right.

""" with gr.Column(visible=False) as flagging_form: flag_txt = gr.Textbox( lines=1, placeholder="Type here...", label="""If you choose to flag your search, we will save the query, language and the number of results you requested. Please consider adding relevant additional context below:""", ) flag_btn = gr.Button("Flag Results") flag_btn.click(flag, inputs=[query, lang, k, flag_txt], outputs=[flag_txt]) def submit(query, lang, k, exact_search): print("submitting", query, lang, k, exact_search) query = query.strip() if query is None or query == "": return "", "" if exact_search: return { results: perform_exact_search(query, k), flagging_form: gr.update(visible=True), } return { results: scisearch(query, lang, k, exact_search), flagging_form: gr.update(visible=True), } query.submit( fn=submit, inputs=[query, lang, k, exact_search], outputs=[results, flagging_form], ) submit_btn.click( submit, inputs=[query, lang, k, exact_search], outputs=[results, flagging_form], ) demo.launch(enable_queue=False, debug=True)