Skip to content

Refine

Init file.

Refine #

Bases: BaseSynthesizer

Refine a response to a query across text chunks.

Source code in llama-index-core/llama_index/core/response_synthesizers/refine.py
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
class Refine(BaseSynthesizer):
    """Refine a response to a query across text chunks."""

    def __init__(
        self,
        llm: Optional[LLMPredictorType] = None,
        callback_manager: Optional[CallbackManager] = None,
        prompt_helper: Optional[PromptHelper] = None,
        text_qa_template: Optional[BasePromptTemplate] = None,
        refine_template: Optional[BasePromptTemplate] = None,
        output_cls: Optional[BaseModel] = None,
        streaming: bool = False,
        verbose: bool = False,
        structured_answer_filtering: bool = False,
        program_factory: Optional[
            Callable[[BasePromptTemplate], BasePydanticProgram]
        ] = None,
        # deprecated
        service_context: Optional[ServiceContext] = None,
    ) -> None:
        if service_context is not None:
            prompt_helper = service_context.prompt_helper

        super().__init__(
            llm=llm,
            callback_manager=callback_manager,
            prompt_helper=prompt_helper,
            service_context=service_context,
            streaming=streaming,
        )
        self._text_qa_template = text_qa_template or DEFAULT_TEXT_QA_PROMPT_SEL
        self._refine_template = refine_template or DEFAULT_REFINE_PROMPT_SEL
        self._verbose = verbose
        self._structured_answer_filtering = structured_answer_filtering
        self._output_cls = output_cls

        if self._streaming and self._structured_answer_filtering:
            raise ValueError(
                "Streaming not supported with structured answer filtering."
            )
        if not self._structured_answer_filtering and program_factory is not None:
            raise ValueError(
                "Program factory not supported without structured answer filtering."
            )
        self._program_factory = program_factory or self._default_program_factory

    def _get_prompts(self) -> PromptDictType:
        """Get prompts."""
        return {
            "text_qa_template": self._text_qa_template,
            "refine_template": self._refine_template,
        }

    def _update_prompts(self, prompts: PromptDictType) -> None:
        """Update prompts."""
        if "text_qa_template" in prompts:
            self._text_qa_template = prompts["text_qa_template"]
        if "refine_template" in prompts:
            self._refine_template = prompts["refine_template"]

    @dispatcher.span
    def get_response(
        self,
        query_str: str,
        text_chunks: Sequence[str],
        prev_response: Optional[RESPONSE_TEXT_TYPE] = None,
        **response_kwargs: Any,
    ) -> RESPONSE_TEXT_TYPE:
        """Give response over chunks."""
        dispatch_event = dispatcher.get_dispatch_event()

        dispatch_event(GetResponseStartEvent())
        response: Optional[RESPONSE_TEXT_TYPE] = None
        for text_chunk in text_chunks:
            if prev_response is None:
                # if this is the first chunk, and text chunk already
                # is an answer, then return it
                response = self._give_response_single(
                    query_str, text_chunk, **response_kwargs
                )
            else:
                # refine response if possible
                response = self._refine_response_single(
                    prev_response, query_str, text_chunk, **response_kwargs
                )
            prev_response = response
        if isinstance(response, str):
            if self._output_cls is not None:
                response = self._output_cls.parse_raw(response)
            else:
                response = response or "Empty Response"
        else:
            response = cast(Generator, response)
        dispatch_event(GetResponseEndEvent())
        return response

    def _default_program_factory(self, prompt: PromptTemplate) -> BasePydanticProgram:
        if self._structured_answer_filtering:
            from llama_index.core.program.utils import get_program_for_llm

            return get_program_for_llm(
                StructuredRefineResponse,
                prompt,
                self._llm,
                verbose=self._verbose,
            )
        else:
            return DefaultRefineProgram(
                prompt=prompt,
                llm=self._llm,
                output_cls=self._output_cls,
            )

    def _give_response_single(
        self,
        query_str: str,
        text_chunk: str,
        **response_kwargs: Any,
    ) -> RESPONSE_TEXT_TYPE:
        """Give response given a query and a corresponding text chunk."""
        text_qa_template = self._text_qa_template.partial_format(query_str=query_str)
        text_chunks = self._prompt_helper.repack(text_qa_template, [text_chunk])

        response: Optional[RESPONSE_TEXT_TYPE] = None
        program = self._program_factory(text_qa_template)
        # TODO: consolidate with loop in get_response_default
        for cur_text_chunk in text_chunks:
            query_satisfied = False
            if response is None and not self._streaming:
                try:
                    structured_response = cast(
                        StructuredRefineResponse,
                        program(
                            context_str=cur_text_chunk,
                            **response_kwargs,
                        ),
                    )
                    query_satisfied = structured_response.query_satisfied
                    if query_satisfied:
                        response = structured_response.answer
                except ValidationError as e:
                    logger.warning(
                        f"Validation error on structured response: {e}", exc_info=True
                    )
            elif response is None and self._streaming:
                response = self._llm.stream(
                    text_qa_template,
                    context_str=cur_text_chunk,
                    **response_kwargs,
                )
                query_satisfied = True
            else:
                response = self._refine_response_single(
                    cast(RESPONSE_TEXT_TYPE, response),
                    query_str,
                    cur_text_chunk,
                    **response_kwargs,
                )
        if response is None:
            response = "Empty Response"
        if isinstance(response, str):
            response = response or "Empty Response"
        else:
            response = cast(Generator, response)
        return response

    def _refine_response_single(
        self,
        response: RESPONSE_TEXT_TYPE,
        query_str: str,
        text_chunk: str,
        **response_kwargs: Any,
    ) -> Optional[RESPONSE_TEXT_TYPE]:
        """Refine response."""
        # TODO: consolidate with logic in response/schema.py
        if isinstance(response, Generator):
            response = get_response_text(response)

        fmt_text_chunk = truncate_text(text_chunk, 50)
        logger.debug(f"> Refine context: {fmt_text_chunk}")
        if self._verbose:
            print(f"> Refine context: {fmt_text_chunk}")

        # NOTE: partial format refine template with query_str and existing_answer here
        refine_template = self._refine_template.partial_format(
            query_str=query_str, existing_answer=response
        )

        # compute available chunk size to see if there is any available space
        # determine if the refine template is too big (which can happen if
        # prompt template + query + existing answer is too large)
        avail_chunk_size = self._prompt_helper._get_available_chunk_size(
            refine_template
        )

        if avail_chunk_size < 0:
            # if the available chunk size is negative, then the refine template
            # is too big and we just return the original response
            return response

        # obtain text chunks to add to the refine template
        text_chunks = self._prompt_helper.repack(
            refine_template, text_chunks=[text_chunk]
        )

        program = self._program_factory(refine_template)
        for cur_text_chunk in text_chunks:
            query_satisfied = False
            if not self._streaming:
                try:
                    structured_response = cast(
                        StructuredRefineResponse,
                        program(
                            context_msg=cur_text_chunk,
                            **response_kwargs,
                        ),
                    )
                    query_satisfied = structured_response.query_satisfied
                    if query_satisfied:
                        response = structured_response.answer
                except ValidationError as e:
                    logger.warning(
                        f"Validation error on structured response: {e}", exc_info=True
                    )
            else:
                # TODO: structured response not supported for streaming
                if isinstance(response, Generator):
                    response = "".join(response)

                refine_template = self._refine_template.partial_format(
                    query_str=query_str, existing_answer=response
                )

                response = self._llm.stream(
                    refine_template,
                    context_msg=cur_text_chunk,
                    **response_kwargs,
                )

        return response

    @dispatcher.span
    async def aget_response(
        self,
        query_str: str,
        text_chunks: Sequence[str],
        prev_response: Optional[RESPONSE_TEXT_TYPE] = None,
        **response_kwargs: Any,
    ) -> RESPONSE_TEXT_TYPE:
        dispatch_event = dispatcher.get_dispatch_event()

        dispatch_event(GetResponseStartEvent())
        response: Optional[RESPONSE_TEXT_TYPE] = None
        for text_chunk in text_chunks:
            if prev_response is None:
                # if this is the first chunk, and text chunk already
                # is an answer, then return it
                response = await self._agive_response_single(
                    query_str, text_chunk, **response_kwargs
                )
            else:
                response = await self._arefine_response_single(
                    prev_response, query_str, text_chunk, **response_kwargs
                )
            prev_response = response
        if response is None:
            response = "Empty Response"
        if isinstance(response, str):
            if self._output_cls is not None:
                response = self._output_cls.parse_raw(response)
            else:
                response = response or "Empty Response"
        else:
            response = cast(AsyncGenerator, response)
        dispatch_event(GetResponseEndEvent())
        return response

    async def _arefine_response_single(
        self,
        response: RESPONSE_TEXT_TYPE,
        query_str: str,
        text_chunk: str,
        **response_kwargs: Any,
    ) -> Optional[RESPONSE_TEXT_TYPE]:
        """Refine response."""
        # TODO: consolidate with logic in response/schema.py
        if isinstance(response, Generator):
            response = get_response_text(response)

        fmt_text_chunk = truncate_text(text_chunk, 50)
        logger.debug(f"> Refine context: {fmt_text_chunk}")

        # NOTE: partial format refine template with query_str and existing_answer here
        refine_template = self._refine_template.partial_format(
            query_str=query_str, existing_answer=response
        )

        # compute available chunk size to see if there is any available space
        # determine if the refine template is too big (which can happen if
        # prompt template + query + existing answer is too large)
        avail_chunk_size = self._prompt_helper._get_available_chunk_size(
            refine_template
        )

        if avail_chunk_size < 0:
            # if the available chunk size is negative, then the refine template
            # is too big and we just return the original response
            return response

        # obtain text chunks to add to the refine template
        text_chunks = self._prompt_helper.repack(
            refine_template, text_chunks=[text_chunk]
        )

        program = self._program_factory(refine_template)
        for cur_text_chunk in text_chunks:
            query_satisfied = False
            if not self._streaming:
                try:
                    structured_response = await program.acall(
                        context_msg=cur_text_chunk,
                        **response_kwargs,
                    )
                    structured_response = cast(
                        StructuredRefineResponse, structured_response
                    )
                    query_satisfied = structured_response.query_satisfied
                    if query_satisfied:
                        response = structured_response.answer
                except ValidationError as e:
                    logger.warning(
                        f"Validation error on structured response: {e}", exc_info=True
                    )
            else:
                if isinstance(response, Generator):
                    response = "".join(response)

                if isinstance(response, AsyncGenerator):
                    _r = ""
                    async for text in response:
                        _r += text
                    response = _r

                refine_template = self._refine_template.partial_format(
                    query_str=query_str, existing_answer=response
                )

                response = await self._llm.astream(
                    refine_template,
                    context_msg=cur_text_chunk,
                    **response_kwargs,
                )

            if query_satisfied:
                refine_template = self._refine_template.partial_format(
                    query_str=query_str, existing_answer=response
                )

        return response

    async def _agive_response_single(
        self,
        query_str: str,
        text_chunk: str,
        **response_kwargs: Any,
    ) -> RESPONSE_TEXT_TYPE:
        """Give response given a query and a corresponding text chunk."""
        text_qa_template = self._text_qa_template.partial_format(query_str=query_str)
        text_chunks = self._prompt_helper.repack(text_qa_template, [text_chunk])

        response: Optional[RESPONSE_TEXT_TYPE] = None
        program = self._program_factory(text_qa_template)
        # TODO: consolidate with loop in get_response_default
        for cur_text_chunk in text_chunks:
            if response is None and not self._streaming:
                try:
                    structured_response = await program.acall(
                        context_str=cur_text_chunk,
                        **response_kwargs,
                    )
                    structured_response = cast(
                        StructuredRefineResponse, structured_response
                    )
                    query_satisfied = structured_response.query_satisfied
                    if query_satisfied:
                        response = structured_response.answer
                except ValidationError as e:
                    logger.warning(
                        f"Validation error on structured response: {e}", exc_info=True
                    )
            elif response is None and self._streaming:
                response = await self._llm.astream(
                    text_qa_template,
                    context_str=cur_text_chunk,
                    **response_kwargs,
                )
                query_satisfied = True
            else:
                response = await self._arefine_response_single(
                    cast(RESPONSE_TEXT_TYPE, response),
                    query_str,
                    cur_text_chunk,
                    **response_kwargs,
                )
        if response is None:
            response = "Empty Response"
        if isinstance(response, str):
            response = response or "Empty Response"
        else:
            response = cast(AsyncGenerator, response)
        return response

get_response #

get_response(query_str: str, text_chunks: Sequence[str], prev_response: Optional[RESPONSE_TEXT_TYPE] = None, **response_kwargs: Any) -> RESPONSE_TEXT_TYPE

Give response over chunks.

Source code in llama-index-core/llama_index/core/response_synthesizers/refine.py
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
@dispatcher.span
def get_response(
    self,
    query_str: str,
    text_chunks: Sequence[str],
    prev_response: Optional[RESPONSE_TEXT_TYPE] = None,
    **response_kwargs: Any,
) -> RESPONSE_TEXT_TYPE:
    """Give response over chunks."""
    dispatch_event = dispatcher.get_dispatch_event()

    dispatch_event(GetResponseStartEvent())
    response: Optional[RESPONSE_TEXT_TYPE] = None
    for text_chunk in text_chunks:
        if prev_response is None:
            # if this is the first chunk, and text chunk already
            # is an answer, then return it
            response = self._give_response_single(
                query_str, text_chunk, **response_kwargs
            )
        else:
            # refine response if possible
            response = self._refine_response_single(
                prev_response, query_str, text_chunk, **response_kwargs
            )
        prev_response = response
    if isinstance(response, str):
        if self._output_cls is not None:
            response = self._output_cls.parse_raw(response)
        else:
            response = response or "Empty Response"
    else:
        response = cast(Generator, response)
    dispatch_event(GetResponseEndEvent())
    return response