跳至内容

FLARE

FLAREInstruct查询引擎 #

基类: BaseQueryEngine

FLARE Instruct 查询引擎。

这是使用鼓励检索指令的FLARE版本。

注意:这是一个测试版功能。接口可能会变更,且可能并不总是给出正确答案。

参数:

名称 类型 描述 默认值
query_engine BaseQueryEngine

要使用的查询引擎

required
llm Optional[LLM]

LLM模型。默认为None。

None
instruct_prompt Optional[PromptTemplate]

指导提示。默认为None。

None
lookahead_answer_inserter Optional[BaseLookaheadAnswerInserter]

前瞻答案插入器。默认为None。

None
done_output_parser Optional[IsDoneOutputParser]

完成输出解析器。 默认为None。

None
query_task_output_parser Optional[QueryTaskOutputParser]

查询任务输出解析器。默认为None。

None
max_iterations int

最大迭代次数。默认为10。

10
max_lookahead_query_tasks int

最大前瞻查询任务数。默认为1。

1
callback_manager Optional[CallbackManager]

回调管理器。 默认为None。

None
verbose bool

提供详细输出。默认为 False。

False
Source code in llama-index-core/llama_index/core/query_engine/flare/base.py
 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
class FLAREInstructQueryEngine(BaseQueryEngine):
    """
    FLARE Instruct query engine.

    This is the version of FLARE that uses retrieval-encouraging instructions.

    NOTE: this is a beta feature. Interfaces might change, and it might not
    always give correct answers.

    Args:
        query_engine (BaseQueryEngine): query engine to use
        llm (Optional[LLM]): LLM model. Defaults to None.
        instruct_prompt (Optional[PromptTemplate]): instruct prompt. Defaults to None.
        lookahead_answer_inserter (Optional[BaseLookaheadAnswerInserter]):
            lookahead answer inserter. Defaults to None.
        done_output_parser (Optional[IsDoneOutputParser]): done output parser.
            Defaults to None.
        query_task_output_parser (Optional[QueryTaskOutputParser]):
            query task output parser. Defaults to None.
        max_iterations (int): max iterations. Defaults to 10.
        max_lookahead_query_tasks (int): max lookahead query tasks. Defaults to 1.
        callback_manager (Optional[CallbackManager]): callback manager.
            Defaults to None.
        verbose (bool): give verbose outputs. Defaults to False.

    """

    def __init__(
        self,
        query_engine: BaseQueryEngine,
        llm: Optional[LLM] = None,
        instruct_prompt: Optional[BasePromptTemplate] = None,
        lookahead_answer_inserter: Optional[BaseLookaheadAnswerInserter] = None,
        done_output_parser: Optional[IsDoneOutputParser] = None,
        query_task_output_parser: Optional[QueryTaskOutputParser] = None,
        max_iterations: int = 10,
        max_lookahead_query_tasks: int = 1,
        callback_manager: Optional[CallbackManager] = None,
        verbose: bool = False,
    ) -> None:
        """Init params."""
        super().__init__(callback_manager=callback_manager)
        self._query_engine = query_engine
        self._llm = llm or Settings.llm
        self._instruct_prompt = instruct_prompt or DEFAULT_INSTRUCT_PROMPT
        self._lookahead_answer_inserter = lookahead_answer_inserter or (
            LLMLookaheadAnswerInserter(llm=self._llm)
        )
        self._done_output_parser = done_output_parser or IsDoneOutputParser()
        self._query_task_output_parser = (
            query_task_output_parser or QueryTaskOutputParser()
        )
        self._max_iterations = max_iterations
        self._max_lookahead_query_tasks = max_lookahead_query_tasks
        self._verbose = verbose

    def _get_prompts(self) -> Dict[str, Any]:
        """Get prompts."""
        return {
            "instruct_prompt": self._instruct_prompt,
        }

    def _update_prompts(self, prompts: PromptDictType) -> None:
        """Update prompts."""
        if "instruct_prompt" in prompts:
            self._instruct_prompt = prompts["instruct_prompt"]

    def _get_prompt_modules(self) -> PromptMixinType:
        """Get prompt sub-modules."""
        return {
            "query_engine": self._query_engine,
            "lookahead_answer_inserter": self._lookahead_answer_inserter,
        }

    def _get_relevant_lookahead_response(self, updated_lookahead_resp: str) -> str:
        """Get relevant lookahead response."""
        # if there's remaining query tasks, then truncate the response
        # until the start position of the first tag
        # there may be remaining query tasks because the _max_lookahead_query_tasks
        # is less than the total number of generated [Search(query)] tags
        remaining_query_tasks = self._query_task_output_parser.parse(
            updated_lookahead_resp
        )
        if len(remaining_query_tasks) == 0:
            relevant_lookahead_resp = updated_lookahead_resp
        else:
            first_task = remaining_query_tasks[0]
            relevant_lookahead_resp = updated_lookahead_resp[: first_task.start_idx]
        return relevant_lookahead_resp

    def _query(self, query_bundle: QueryBundle) -> RESPONSE_TYPE:
        """Query and get response."""
        print_text(f"Query: {query_bundle.query_str}\n", color="green")
        cur_response = ""
        source_nodes = []
        for iter in range(self._max_iterations):
            if self._verbose:
                print_text(f"Current response: {cur_response}\n", color="blue")
            # generate "lookahead response" that contains "[Search(query)]" tags
            # e.g.
            # The colors on the flag of Ghana have the following meanings. Red is
            # for [Search(Ghana flag meaning)],...
            lookahead_resp = self._llm.predict(
                self._instruct_prompt,
                query_str=query_bundle.query_str,
                existing_answer=cur_response,
            )
            lookahead_resp = lookahead_resp.strip()
            if self._verbose:
                print_text(f"Lookahead response: {lookahead_resp}\n", color="pink")

            is_done, fmt_lookahead = self._done_output_parser.parse(lookahead_resp)
            if is_done:
                cur_response = cur_response.strip() + " " + fmt_lookahead.strip()
                break

            # parse lookahead response into query tasks
            query_tasks = self._query_task_output_parser.parse(lookahead_resp)

            # get answers for each query task
            query_tasks = query_tasks[: self._max_lookahead_query_tasks]
            query_answers = []
            for _, query_task in enumerate(query_tasks):
                answer_obj = self._query_engine.query(query_task.query_str)
                if not isinstance(answer_obj, Response):
                    raise ValueError(
                        f"Expected Response object, got {type(answer_obj)} instead."
                    )
                query_answer = str(answer_obj)
                query_answers.append(query_answer)
                source_nodes.extend(answer_obj.source_nodes)

            # fill in the lookahead response template with the query answers
            # from the query engine
            updated_lookahead_resp = self._lookahead_answer_inserter.insert(
                lookahead_resp, query_tasks, query_answers, prev_response=cur_response
            )

            # get "relevant" lookahead response by truncating the updated
            # lookahead response until the start position of the first tag
            # also remove the prefix from the lookahead response, so that
            # we can concatenate it with the existing response
            relevant_lookahead_resp_wo_prefix = self._get_relevant_lookahead_response(
                updated_lookahead_resp
            )

            if self._verbose:
                print_text(
                    "Updated lookahead response: "
                    + f"{relevant_lookahead_resp_wo_prefix}\n",
                    color="pink",
                )

            # append the relevant lookahead response to the final response
            cur_response = (
                cur_response.strip() + " " + relevant_lookahead_resp_wo_prefix.strip()
            )

        # NOTE: at the moment, does not support streaming
        return Response(response=cur_response, source_nodes=source_nodes)

    async def _aquery(self, query_bundle: QueryBundle) -> RESPONSE_TYPE:
        return self._query(query_bundle)

    def retrieve(self, query_bundle: QueryBundle) -> List[NodeWithScore]:
        # if the query engine is a retriever, then use the retrieve method
        if isinstance(self._query_engine, RetrieverQueryEngine):
            return self._query_engine.retrieve(query_bundle)
        else:
            raise NotImplementedError(
                "This query engine does not support retrieve, use query directly"
            )

    async def aretrieve(self, query_bundle: QueryBundle) -> List[NodeWithScore]:
        # if the query engine is a retriever, then use the retrieve method
        if isinstance(self._query_engine, RetrieverQueryEngine):
            return await self._query_engine.aretrieve(query_bundle)
        else:
            raise NotImplementedError(
                "This query engine does not support retrieve, use query directly"
            )