api/db/services/dialog_service.py

#
#  Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
#
import logging
import binascii
import os
import json
import re
from collections import defaultdict
from copy import deepcopy
from timeit import default_timer as timer
import datetime
from datetime import timedelta
from api.db import LLMType, ParserType,StatusEnum
from api.db.db_models import Dialog, DB
from api.db.services.common_service import CommonService
from api.db.services.knowledgebase_service import KnowledgebaseService
from api.db.services.llm_service import LLMService, TenantLLMService, LLMBundle
from api import settings
from rag.app.resume import forbidden_select_fields4resume
from rag.nlp.search import index_name
from rag.utils import rmSpace, num_tokens_from_string, encoder
from api.utils.file_utils import get_project_base_directory


class DialogService(CommonService):
    model = Dialog

    @classmethod
    @DB.connection_context()
    def get_list(cls, tenant_id,
                 page_number, items_per_page, orderby, desc, id , name):
        chats = cls.model.select()
        if id:
            chats = chats.where(cls.model.id == id)
        if name:
            chats = chats.where(cls.model.name == name)
        chats = chats.where(
              (cls.model.tenant_id == tenant_id)
            & (cls.model.status == StatusEnum.VALID.value)
        )
        if desc:
            chats = chats.order_by(cls.model.getter_by(orderby).desc())
        else:
            chats = chats.order_by(cls.model.getter_by(orderby).asc())

        chats = chats.paginate(page_number, items_per_page)

        return list(chats.dicts())


def message_fit_in(msg, max_length=4000):
    def count():
        nonlocal msg
        tks_cnts = []
        for m in msg:
            tks_cnts.append(
                {"role": m["role"], "count": num_tokens_from_string(m["content"])})
        total = 0
        for m in tks_cnts:
            total += m["count"]
        return total

    c = count()
    if c < max_length:
        return c, msg

    msg_ = [m for m in msg[:-1] if m["role"] == "system"]
    if len(msg) > 1:
        msg_.append(msg[-1])
    msg = msg_
    c = count()
    if c < max_length:
        return c, msg

    ll = num_tokens_from_string(msg_[0]["content"])
    ll2 = num_tokens_from_string(msg_[-1]["content"])
    if ll / (ll + ll2) > 0.8:
        m = msg_[0]["content"]
        m = encoder.decode(encoder.encode(m)[:max_length - ll2])
        msg[0]["content"] = m
        return max_length, msg

    m = msg_[1]["content"]
    m = encoder.decode(encoder.encode(m)[:max_length - ll2])
    msg[1]["content"] = m
    return max_length, msg


def llm_id2llm_type(llm_id):
    llm_id, _ = TenantLLMService.split_model_name_and_factory(llm_id)
    fnm = os.path.join(get_project_base_directory(), "conf")
    llm_factories = json.load(open(os.path.join(fnm, "llm_factories.json"), "r"))
    for llm_factory in llm_factories["factory_llm_infos"]:
        for llm in llm_factory["llm"]:
            if llm_id == llm["llm_name"]:
                return llm["model_type"].strip(",")[-1]


def kb_prompt(kbinfos, max_tokens):
    knowledges = [ck["content_with_weight"] for ck in kbinfos["chunks"]]
    used_token_count = 0
    chunks_num = 0
    for i, c in enumerate(knowledges):
        used_token_count += num_tokens_from_string(c)
        chunks_num += 1
        if max_tokens * 0.97 < used_token_count:
            knowledges = knowledges[:i]
            break

    doc2chunks = defaultdict(list)
    for i, ck in enumerate(kbinfos["chunks"]):
        if i >= chunks_num:
            break
        doc2chunks["docnm_kwd"].append(ck["content_with_weight"])

    knowledges = []
    for nm, chunks in doc2chunks.items():
        txt = f"Document: {nm} \nContains the following relevant fragments:\n"
        for i, chunk in enumerate(chunks, 1):
            txt += f"{i}. {chunk}\n"
        knowledges.append(txt)
    return knowledges


def chat(dialog, messages, stream=True, **kwargs):
    assert messages[-1]["role"] == "user", "The last content of this conversation is not from user."
    st = timer()
    llm_id, fid = TenantLLMService.split_model_name_and_factory(dialog.llm_id)
    llm = LLMService.query(llm_name=llm_id) if not fid else LLMService.query(llm_name=llm_id, fid=fid)
    if not llm:
        llm = TenantLLMService.query(tenant_id=dialog.tenant_id, llm_name=llm_id) if not fid else \
            TenantLLMService.query(tenant_id=dialog.tenant_id, llm_name=llm_id, llm_factory=fid)
        if not llm:
            raise LookupError("LLM(%s) not found" % dialog.llm_id)
        max_tokens = 8192
    else:
        max_tokens = llm[0].max_tokens
    kbs = KnowledgebaseService.get_by_ids(dialog.kb_ids)
    embd_nms = list(set([kb.embd_id for kb in kbs]))
    if len(embd_nms) != 1:
        yield {"answer": "**ERROR**: Knowledge bases use different embedding models.", "reference": []}
        return {"answer": "**ERROR**: Knowledge bases use different embedding models.", "reference": []}

    is_kg = all([kb.parser_id == ParserType.KG for kb in kbs])
    retr = settings.retrievaler if not is_kg else settings.kg_retrievaler

    questions = [m["content"] for m in messages if m["role"] == "user"][-3:]
    attachments = kwargs["doc_ids"].split(",") if "doc_ids" in kwargs else None
    if "doc_ids" in messages[-1]:
        attachments = messages[-1]["doc_ids"]
        for m in messages[:-1]:
            if "doc_ids" in m:
                attachments.extend(m["doc_ids"])

    embd_mdl = LLMBundle(dialog.tenant_id, LLMType.EMBEDDING, embd_nms[0])
    if not embd_mdl:
        raise LookupError("Embedding model(%s) not found" % embd_nms[0])

    if llm_id2llm_type(dialog.llm_id) == "image2text":
        chat_mdl = LLMBundle(dialog.tenant_id, LLMType.IMAGE2TEXT, dialog.llm_id)
    else:
        chat_mdl = LLMBundle(dialog.tenant_id, LLMType.CHAT, dialog.llm_id)

    prompt_config = dialog.prompt_config
    field_map = KnowledgebaseService.get_field_map(dialog.kb_ids)
    tts_mdl = None
    if prompt_config.get("tts"):
        tts_mdl = LLMBundle(dialog.tenant_id, LLMType.TTS)
    # try to use sql if field mapping is good to go
    if field_map:
        logging.debug("Use SQL to retrieval:{}".format(questions[-1]))
        ans = use_sql(questions[-1], field_map, dialog.tenant_id, chat_mdl, prompt_config.get("quote", True))
        if ans:
            yield ans
            return

    for p in prompt_config["parameters"]:
        if p["key"] == "knowledge":
            continue
        if p["key"] not in kwargs and not p["optional"]:
            raise KeyError("Miss parameter: " + p["key"])
        if p["key"] not in kwargs:
            prompt_config["system"] = prompt_config["system"].replace(
                "{%s}" % p["key"], " ")

    if len(questions) > 1 and prompt_config.get("refine_multiturn"):
        questions = [full_question(dialog.tenant_id, dialog.llm_id, messages)]
    else:
        questions = questions[-1:]
    refineQ_tm = timer()
    keyword_tm = timer()

    rerank_mdl = None
    if dialog.rerank_id:
        rerank_mdl = LLMBundle(dialog.tenant_id, LLMType.RERANK, dialog.rerank_id)

    for _ in range(len(questions) // 2):
        questions.append(questions[-1])
    if "knowledge" not in [p["key"] for p in prompt_config["parameters"]]:
        kbinfos = {"total": 0, "chunks": [], "doc_aggs": []}
    else:
        if prompt_config.get("keyword", False):
            questions[-1] += keyword_extraction(chat_mdl, questions[-1])
            keyword_tm = timer()

        tenant_ids = list(set([kb.tenant_id for kb in kbs]))
        kbinfos = retr.retrieval(" ".join(questions), embd_mdl, tenant_ids, dialog.kb_ids, 1, dialog.top_n,
                                        dialog.similarity_threshold,
                                        dialog.vector_similarity_weight,
                                        doc_ids=attachments,
                                        top=dialog.top_k, aggs=False, rerank_mdl=rerank_mdl)
    knowledges = kb_prompt(kbinfos, max_tokens)
    logging.debug(
        "{}->{}".format(" ".join(questions), "\n->".join(knowledges)))
    retrieval_tm = timer()

    if not knowledges and prompt_config.get("empty_response"):
        empty_res = prompt_config["empty_response"]
        yield {"answer": empty_res, "reference": kbinfos, "audio_binary": tts(tts_mdl, empty_res)}
        return {"answer": prompt_config["empty_response"], "reference": kbinfos}

    kwargs["knowledge"] = "\n\n------\n\n".join(knowledges)
    gen_conf = dialog.llm_setting

    msg = [{"role": "system", "content": prompt_config["system"].format(**kwargs)}]
    msg.extend([{"role": m["role"], "content": re.sub(r"##\d+\$\$", "", m["content"])}
                for m in messages if m["role"] != "system"])
    used_token_count, msg = message_fit_in(msg, int(max_tokens * 0.97))
    assert len(msg) >= 2, f"message_fit_in has bug: {msg}"
    prompt = msg[0]["content"]
    prompt += "\n\n### Query:\n%s" % " ".join(questions)

    if "max_tokens" in gen_conf:
        gen_conf["max_tokens"] = min(
            gen_conf["max_tokens"],
            max_tokens - used_token_count)

    def decorate_answer(answer):
        nonlocal prompt_config, knowledges, kwargs, kbinfos, prompt, retrieval_tm
        refs = []
        if knowledges and (prompt_config.get("quote", True) and kwargs.get("quote", True)):
            answer, idx = retr.insert_citations(answer,
                                                       [ck["content_ltks"]
                                                        for ck in kbinfos["chunks"]],
                                                       [ck["vector"]
                                                        for ck in kbinfos["chunks"]],
                                                       embd_mdl,
                                                       tkweight=1 - dialog.vector_similarity_weight,
                                                       vtweight=dialog.vector_similarity_weight)
            idx = set([kbinfos["chunks"][int(i)]["doc_id"] for i in idx])
            recall_docs = [
                d for d in kbinfos["doc_aggs"] if d["doc_id"] in idx]
            if not recall_docs:
                recall_docs = kbinfos["doc_aggs"]
            kbinfos["doc_aggs"] = recall_docs

            refs = deepcopy(kbinfos)
            for c in refs["chunks"]:
                if c.get("vector"):
                    del c["vector"]

        if answer.lower().find("invalid key") >= 0 or answer.lower().find("invalid api") >= 0:
            answer += " Please set LLM API-Key in 'User Setting -> Model providers -> API-Key'"
        done_tm = timer()
        prompt += "\n\n### Elapsed\n  - Refine Question: %.1f ms\n  - Keywords: %.1f ms\n  - Retrieval: %.1f ms\n  - LLM: %.1f ms" % (
            (refineQ_tm - st) * 1000, (keyword_tm - refineQ_tm) * 1000, (retrieval_tm - keyword_tm) * 1000,
            (done_tm - retrieval_tm) * 1000)
        return {"answer": answer, "reference": refs, "prompt": prompt}

    if stream:
        last_ans = ""
        answer = ""
        for ans in chat_mdl.chat_streamly(prompt, msg[1:], gen_conf):
            answer = ans
            delta_ans = ans[len(last_ans):]
            if num_tokens_from_string(delta_ans) < 16:
                continue
            last_ans = answer
            yield {"answer": answer, "reference": {}, "audio_binary": tts(tts_mdl, delta_ans)}
        delta_ans = answer[len(last_ans):]
        if delta_ans:
            yield {"answer": answer, "reference": {}, "audio_binary": tts(tts_mdl, delta_ans)}
        yield decorate_answer(answer)
    else:
        answer = chat_mdl.chat(prompt, msg[1:], gen_conf)
        logging.debug("User: {}|Assistant: {}".format(
            msg[-1]["content"], answer))
        res = decorate_answer(answer)
        res["audio_binary"] = tts(tts_mdl, answer)
        yield res


def use_sql(question, field_map, tenant_id, chat_mdl, quota=True):
    sys_prompt = "你是一个DBA。你需要这对以下表的字段结构，根据用户的问题列表，写出最后一个问题对应的SQL。"
    user_promt = """
表名：{}；
数据库表字段说明如下：
{}

问题如下：
{}
请写出SQL, 且只要SQL，不要有其他说明及文字。
""".format(
        index_name(tenant_id),
        "\n".join([f"{k}: {v}" for k, v in field_map.items()]),
        question
    )
    tried_times = 0

    def get_table():
        nonlocal sys_prompt, user_promt, question, tried_times
        sql = chat_mdl.chat(sys_prompt, [{"role": "user", "content": user_promt}], {
            "temperature": 0.06})
        logging.debug(f"{question} ==> {user_promt} get SQL: {sql}")
        sql = re.sub(r"[\r\n]+", " ", sql.lower())
        sql = re.sub(r".*select ", "select ", sql.lower())
        sql = re.sub(r" +", " ", sql)
        sql = re.sub(r"([;；]|```).*", "", sql)
        if sql[:len("select ")] != "select ":
            return None, None
        if not re.search(r"((sum|avg|max|min)\(|group by )", sql.lower()):
            if sql[:len("select *")] != "select *":
                sql = "select doc_id,docnm_kwd," + sql[6:]
            else:
                flds = []
                for k in field_map.keys():
                    if k in forbidden_select_fields4resume:
                        continue
                    if len(flds) > 11:
                        break
                    flds.append(k)
                sql = "select doc_id,docnm_kwd," + ",".join(flds) + sql[8:]

        logging.debug(f"{question} get SQL(refined): {sql}")
        tried_times += 1
        return settings.retrievaler.sql_retrieval(sql, format="json"), sql

    tbl, sql = get_table()
    if tbl is None:
        return None
    if tbl.get("error") and tried_times <= 2:
        user_promt = """
        表名：{}；
        数据库表字段说明如下：
        {}

        问题如下：
        {}

        你上一次给出的错误SQL如下：
        {}

        后台报错如下：
        {}

        请纠正SQL中的错误再写一遍，且只要SQL，不要有其他说明及文字。
        """.format(
            index_name(tenant_id),
            "\n".join([f"{k}: {v}" for k, v in field_map.items()]),
            question, sql, tbl["error"]
        )
        tbl, sql = get_table()
        logging.debug("TRY it again: {}".format(sql))

    logging.debug("GET table: {}".format(tbl))
    if tbl.get("error") or len(tbl["rows"]) == 0:
        return None

    docid_idx = set([ii for ii, c in enumerate(
        tbl["columns"]) if c["name"] == "doc_id"])
    docnm_idx = set([ii for ii, c in enumerate(
        tbl["columns"]) if c["name"] == "docnm_kwd"])
    clmn_idx = [ii for ii in range(
        len(tbl["columns"])) if ii not in (docid_idx | docnm_idx)]

    # compose markdown table
    clmns = "|" + "|".join([re.sub(r"(/.*|（[^（）]+）)", "", field_map.get(tbl["columns"][i]["name"],
                                                                        tbl["columns"][i]["name"])) for i in
                            clmn_idx]) + ("|Source|" if docid_idx and docid_idx else "|")

    line = "|" + "|".join(["------" for _ in range(len(clmn_idx))]) + \
           ("|------|" if docid_idx and docid_idx else "")

    rows = ["|" +
            "|".join([rmSpace(str(r[i])) for i in clmn_idx]).replace("None", " ") +
            "|" for r in tbl["rows"]]
    rows = [r for r in rows if re.sub(r"[ |]+", "", r)]
    if quota:
        rows = "\n".join([r + f" ##{ii}$$ |" for ii, r in enumerate(rows)])
    else:
        rows = "\n".join([r + f" ##{ii}$$ |" for ii, r in enumerate(rows)])
    rows = re.sub(r"T[0-9]{2}:[0-9]{2}:[0-9]{2}(\.[0-9]+Z)?\|", "|", rows)

    if not docid_idx or not docnm_idx:
        logging.warning("SQL missing field: " + sql)
        return {
            "answer": "\n".join([clmns, line, rows]),
            "reference": {"chunks": [], "doc_aggs": []},
            "prompt": sys_prompt
        }

    docid_idx = list(docid_idx)[0]
    docnm_idx = list(docnm_idx)[0]
    doc_aggs = {}
    for r in tbl["rows"]:
        if r[docid_idx] not in doc_aggs:
            doc_aggs[r[docid_idx]] = {"doc_name": r[docnm_idx], "count": 0}
        doc_aggs[r[docid_idx]]["count"] += 1
    return {
        "answer": "\n".join([clmns, line, rows]),
        "reference": {"chunks": [{"doc_id": r[docid_idx], "docnm_kwd": r[docnm_idx]} for r in tbl["rows"]],
                      "doc_aggs": [{"doc_id": did, "doc_name": d["doc_name"], "count": d["count"]} for did, d in
                                   doc_aggs.items()]},
        "prompt": sys_prompt
    }


def relevant(tenant_id, llm_id, question, contents: list):
    if llm_id2llm_type(llm_id) == "image2text":
        chat_mdl = LLMBundle(tenant_id, LLMType.IMAGE2TEXT, llm_id)
    else:
        chat_mdl = LLMBundle(tenant_id, LLMType.CHAT, llm_id)
    prompt = """
        You are a grader assessing relevance of a retrieved document to a user question. 
        It does not need to be a stringent test. The goal is to filter out erroneous retrievals.
        If the document contains keyword(s) or semantic meaning related to the user question, grade it as relevant. 
        Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question.
        No other words needed except 'yes' or 'no'.
    """
    if not contents:
        return False
    contents = "Documents: \n" + "   - ".join(contents)
    contents = f"Question: {question}\n" + contents
    if num_tokens_from_string(contents) >= chat_mdl.max_length - 4:
        contents = encoder.decode(encoder.encode(contents)[:chat_mdl.max_length - 4])
    ans = chat_mdl.chat(prompt, [{"role": "user", "content": contents}], {"temperature": 0.01})
    if ans.lower().find("yes") >= 0:
        return True
    return False


def rewrite(tenant_id, llm_id, question):
    if llm_id2llm_type(llm_id) == "image2text":
        chat_mdl = LLMBundle(tenant_id, LLMType.IMAGE2TEXT, llm_id)
    else:
        chat_mdl = LLMBundle(tenant_id, LLMType.CHAT, llm_id)
    prompt = """
        You are an expert at query expansion to generate a paraphrasing of a question.
        I can't retrieval relevant information from the knowledge base by using user's question directly.     
        You need to expand or paraphrase user's question by multiple ways such as using synonyms words/phrase, 
        writing the abbreviation in its entirety, adding some extra descriptions or explanations, 
        changing the way of expression, translating the original question into another language (English/Chinese), etc. 
        And return 5 versions of question and one is from translation.
        Just list the question. No other words are needed.
    """
    ans = chat_mdl.chat(prompt, [{"role": "user", "content": question}], {"temperature": 0.8})
    return ans


def keyword_extraction(chat_mdl, content, topn=3):
    prompt = f"""
Role: You're a text analyzer. 
Task: extract the most important keywords/phrases of a given piece of text content.
Requirements: 
  - Summarize the text content, and give top {topn} important keywords/phrases.
  - The keywords MUST be in language of the given piece of text content.
  - The keywords are delimited by ENGLISH COMMA.
  - Keywords ONLY in output.

### Text Content 
{content}

"""
    msg = [
        {"role": "system", "content": prompt},
        {"role": "user", "content": "Output: "}
    ]
    _, msg = message_fit_in(msg, chat_mdl.max_length)
    kwd = chat_mdl.chat(prompt, msg[1:], {"temperature": 0.2})
    if isinstance(kwd, tuple):
        kwd = kwd[0]
    if kwd.find("**ERROR**") >=0:
        return ""
    return kwd


def question_proposal(chat_mdl, content, topn=3):
    prompt = f"""
Role: You're a text analyzer. 
Task:  propose {topn} questions about a given piece of text content.
Requirements: 
  - Understand and summarize the text content, and propose top {topn} important questions.
  - The questions SHOULD NOT have overlapping meanings.
  - The questions SHOULD cover the main content of the text as much as possible.
  - The questions MUST be in language of the given piece of text content.
  - One question per line.
  - Question ONLY in output.

### Text Content 
{content}

"""
    msg = [
        {"role": "system", "content": prompt},
        {"role": "user", "content": "Output: "}
    ]
    _, msg = message_fit_in(msg, chat_mdl.max_length)
    kwd = chat_mdl.chat(prompt, msg[1:], {"temperature": 0.2})
    if isinstance(kwd, tuple):
        kwd = kwd[0]
    if kwd.find("**ERROR**") >= 0:
        return ""
    return kwd


def full_question(tenant_id, llm_id, messages):
    if llm_id2llm_type(llm_id) == "image2text":
        chat_mdl = LLMBundle(tenant_id, LLMType.IMAGE2TEXT, llm_id)
    else:
        chat_mdl = LLMBundle(tenant_id, LLMType.CHAT, llm_id)
    conv = []
    for m in messages:
        if m["role"] not in ["user", "assistant"]:
            continue
        conv.append("{}: {}".format(m["role"].upper(), m["content"]))
    conv = "\n".join(conv)
    today = datetime.date.today().isoformat()
    yesterday = (datetime.date.today() - timedelta(days=1)).isoformat()
    tomorrow = (datetime.date.today() + timedelta(days=1)).isoformat()
    prompt = f"""
Role: A helpful assistant

Task and steps: 
    1. Generate a full user question that would follow the conversation.
    2. If the user's question involves relative date, you need to convert it into absolute date based on the current date, which is {today}. For example: 'yesterday' would be converted to {yesterday}.
    
Requirements & Restrictions:
  - Text generated MUST be in the same language of the original user's question.
  - If the user's latest question is completely, don't do anything, just return the original question.
  - DON'T generate anything except a refined question.

######################
-Examples-
######################

# Example 1
## Conversation
USER: What is the name of Donald Trump's father?
ASSISTANT:  Fred Trump.
USER: And his mother?
###############
Output: What's the name of Donald Trump's mother?

------------
# Example 2
## Conversation
USER: What is the name of Donald Trump's father?
ASSISTANT:  Fred Trump.
USER: And his mother?
ASSISTANT:  Mary Trump.
User: What's her full name?
###############
Output: What's the full name of Donald Trump's mother Mary Trump?

------------
# Example 3
## Conversation
USER: What's the weather today in London?
ASSISTANT:  Cloudy.
USER: What's about tomorrow in Rochester?
###############
Output: What's the weather in Rochester on {tomorrow}?
######################

# Real Data
## Conversation
{conv}
###############
    """
    ans = chat_mdl.chat(prompt, [{"role": "user", "content": "Output: "}], {"temperature": 0.2})
    return ans if ans.find("**ERROR**") < 0 else messages[-1]["content"]


def tts(tts_mdl, text):
    if not tts_mdl or not text:
        return
    bin = b""
    for chunk in tts_mdl.tts(text):
        bin += chunk
    return binascii.hexlify(bin).decode("utf-8")


def ask(question, kb_ids, tenant_id):
    kbs = KnowledgebaseService.get_by_ids(kb_ids)
    embd_nms = list(set([kb.embd_id for kb in kbs]))

    is_kg = all([kb.parser_id == ParserType.KG for kb in kbs])
    retr = settings.retrievaler if not is_kg else settings.kg_retrievaler

    embd_mdl = LLMBundle(tenant_id, LLMType.EMBEDDING, embd_nms[0])
    chat_mdl = LLMBundle(tenant_id, LLMType.CHAT)
    max_tokens = chat_mdl.max_length
    tenant_ids = list(set([kb.tenant_id for kb in kbs]))
    kbinfos = retr.retrieval(question, embd_mdl, tenant_ids, kb_ids, 1, 12, 0.1, 0.3, aggs=False)
    knowledges = kb_prompt(kbinfos, max_tokens)
    prompt = """
    Role: You're a smart assistant. Your name is Miss R.
    Task: Summarize the information from knowledge bases and answer user's question.
    Requirements and restriction:
      - DO NOT make things up, especially for numbers.
      - If the information from knowledge is irrelevant with user's question, JUST SAY: Sorry, no relevant information provided.
      - Answer with markdown format text.
      - Answer in language of user's question.
      - DO NOT make things up, especially for numbers.

    ### Information from knowledge bases
    %s

    The above is information from knowledge bases.

    """ % "\n".join(knowledges)
    msg = [{"role": "user", "content": question}]

    def decorate_answer(answer):
        nonlocal knowledges, kbinfos, prompt
        answer, idx = retr.insert_citations(answer,
                                            [ck["content_ltks"]
                                             for ck in kbinfos["chunks"]],
                                            [ck["vector"]
                                             for ck in kbinfos["chunks"]],
                                            embd_mdl,
                                            tkweight=0.7,
                                            vtweight=0.3)
        idx = set([kbinfos["chunks"][int(i)]["doc_id"] for i in idx])
        recall_docs = [
            d for d in kbinfos["doc_aggs"] if d["doc_id"] in idx]
        if not recall_docs:
            recall_docs = kbinfos["doc_aggs"]
        kbinfos["doc_aggs"] = recall_docs
        refs = deepcopy(kbinfos)
        for c in refs["chunks"]:
            if c.get("vector"):
                del c["vector"]

        if answer.lower().find("invalid key") >= 0 or answer.lower().find("invalid api") >= 0:
            answer += " Please set LLM API-Key in 'User Setting -> Model Providers -> API-Key'"
        return {"answer": answer, "reference": refs}

    answer = ""
    for ans in chat_mdl.chat_streamly(prompt, msg, {"temperature": 0.1}):
        answer = ans
        yield {"answer": answer, "reference": {}}
    yield decorate_answer(answer)