43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
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
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 | class Querier(Chain):
"""Queriers are responsible for answering questions about the dataset,
using the pre-populated model.
The high level flow is as follows:
1. Determine a set of sub-questions necessary to answer the original.
2. For each sub-question, determine a set of queries that would render relevant facts.
3. For each query, search the vector store for facts _or_ queries that are similar to the embedded query.
4. Extract the facts from these results.
5. Recursively summarize up the tree until the original question is answered.
"""
FACT_PROMPT = PromptTemplate(
input_variables=["fact", "context", "attributes"],
template=dedent(
"""
The user was tagged as: {attributes}.
Their response was: {fact}
A summary of the whole interview is: {context}
"""
),
)
def __init__(self, index: str, debug: bool = False):
super().__init__(debug=debug)
self.index_name = index
self.embeddings = OpenAIEmbeddings()
self.summarizer = Summarizer()
self.index = pinecone.Index(index)
self.facts = set()
# Questions
def steps_template(self):
"""The template to determine sub-questions."""
return PromptTemplate(
template=dedent(
"""
Your task is to determine a set of at most {n} steps which would answer a question.
You must not answer the question, merely determine the best way to answer it.
For example, if the question was "What is the most popular house colors?":
1. Determine all the possible colors that houses can be.
2. Determine the number of houses that are each color.
3. Determine the most popular color.
The question is: {query}
1.
"""
),
input_variables=["query", "n"],
)
def queries_template(self):
"""The template to determine sub-question queries."""
return PromptTemplate(
template=dedent(
"""
Your task is to determine a set of natural-language queries which would answer a question.
The queries run against a knowledgebase of facts compiled across several user interviews.
The overall question you are trying to answer is: {query}
You are on the following step: {step}
Generate a bulleted list of at most {n} natural-language queries to complete this step.
-
"""
),
input_variables=["query", "step", "n"],
)
# Answers
def facts_template(self, facts: list[Fact]):
"""The template to summarize sub-question queries."""
return FewShotPromptTemplate(
examples=cast(list[dict], facts),
example_prompt=self.FACT_PROMPT,
prefix=dedent(
"""
Your task is to answer a query against a corpus of user interviews.
To help answer the question, you are provided with a set of facts (along with the context and attributes of the author of the fact).
If it is not possible to answer, say that you do not know the answer.
The query is: {query}
The relevant facts are:
"""
),
suffix="Your response:\n",
input_variables=["query"],
example_separator="\n",
)
def answers_template(self, answers: list[Answer]):
"""The template to summarize sub-questions."""
return FewShotPromptTemplate(
examples=cast(list[dict], answers),
example_prompt=PromptTemplate(
template=dedent(
"""
Query:
{question}
Answer:
{answer}
"""
),
input_variables=["question", "answer"],
),
prefix=dedent(
"""
Your task is to take a set of queries and answers, and use them to complete a step towards answering an original question.
The original question you are trying to answer is: {query}
You are on the following step: {step}
Here are the queries and answers:
"""
),
suffix="Completed step:\n",
input_variables=["query", "step"],
)
def conclusions_template(self, conclusions: list[Conclusion]):
"""The template to summarize the final answer from a set of conclusions."""
return FewShotPromptTemplate(
examples=cast(list[dict], conclusions),
example_prompt=PromptTemplate(
template=dedent(
"""
Step:
{step}
Conclusion:
{conclusion}
"""
),
input_variables=["step", "conclusion"],
),
prefix=dedent(
"""
Your task is to take a set of steps that were conducted to answer a question, and use them to answer that question.
Answer the question in a structured manner, using the format requested. For example, if the question specifies a list of properties, render a table with that list.
The question you are trying to answer: {query}
The steps you went through to answer this question are:
"""
),
suffix="Final answer:\n",
input_variables=["query"],
)
def structured_data_template(self, metrics: dict):
"""The template to summarize the final answer from the collection of structured data."""
return PromptTemplate(
template=dedent(
f"""
Your task is to take a set of data that was collected about a collection of interviews, and use it to answer a question.
Answer the question in a structured manner, using the format requested. For example, if the question specifies a list of properties, render a table with that list.
The question you are trying to answer: {{{{ query }}}}
The structured data you collected along the way:
{json.dumps(metrics)}
Your answer:
"""
),
input_variables=["query"],
template_format="jinja2",
)
def meta_conclusions_template(self, answers: list[dict]):
"""The template to pick between several final answers."""
return FewShotPromptTemplate(
examples=answers,
example_prompt=PromptTemplate(
template=dedent(
"""
Method:
{method}
Answer:
```
{answer}
```
"""
),
input_variables=["method", "answer"],
),
prefix=dedent(
"""
An answer was procuced for a question using several different methods.
First, evaluate how clear, specific, and thorough each answer is.
Then, select the best one and return it inside a code block.
If you are unsure what the best answer is, use the most precise one.
You can clean up the answer as you return it, but do not change the meaning.
The question is: {query}
"""
),
suffix="Evaluation and Returned Answer:\nEvaluation:\n1.",
input_variables=["query"],
)
@overload
def _query(self, prompt: BasePromptTemplate, **kwargs) -> str:
...
@overload
def _query(
self, prompt: BasePromptTemplate, initial: str, prefix: str, **kwargs
) -> list[str]:
...
@retry(
exceptions=RateLimitError,
tries=5,
delay=10,
backoff=2,
max_delay=120,
jitter=(0, 10),
)
def _query(
self,
prompt: BasePromptTemplate,
initial: str = "",
prefix: str = "",
quiet: bool = False,
**kwargs,
):
chain = LLMChain(llm=self.llm, prompt=prompt)
results = initial + chain.run(**kwargs)
if not quiet:
self.dprint(results)
if initial and prefix:
return self._parse(results.splitlines(), prefix)
else:
return results
def _query_facts(self, query: str, n: int, classes: list[Classes]):
embedding = self.embeddings.embed_query(query)
filter = {"$or": [{"classes": c.value} for c in classes]} if classes else None
results = self.index.query(
embedding, top_k=n * 3, filter=filter # type: ignore
)["matches"]
facts: list[Fact] = [
{
"fact": e.fact,
"context": e.document.metadata["summary"],
"attributes": ", ".join(
itertools.chain.from_iterable(
e.document.metadata["classes"].values()
)
),
}
for r in results
for e in [Embedding.safe_get(r["id"])]
if e
]
new_facts = {f["fact"]: f for f in facts if f["fact"] not in self.facts}
old_facts = {f["fact"]: f for f in facts if f["fact"] in self.facts}
facts = (list(new_facts.values()) + list(old_facts.values()))[:n]
self.facts.update(f["fact"] for f in facts)
if not facts:
raise RuntimeError("No vectors found!")
return facts
def _answer_question(self, question: str, n: int, classes: list[Classes]) -> Answer:
self.dprint(f"Ask", question, color="magenta")
with self.dprint.indent_children():
facts = self._query_facts(question, n, classes)
answer = self.summarizer.summarize_facts(
question,
[Document(page_content=self.FACT_PROMPT.format(**f)) for f in facts],
)
self.dprint("Answer", answer)
return {"question": question, "answer": answer}
def _conclude_step(
self, step: str, query: str, n: int, classes: list[Classes]
) -> Conclusion:
self.dprint("Action", step, color="cyan")
with self.dprint.indent_children():
questions = self._query(
self.queries_template(), "-", r"\-", query=query, step=step, n=n
)
answers = self._pmap(self._answer_question, questions, n, classes)
conclusion = self._query(
self.answers_template(answers), query=query, step=step
)
return {"step": step, "conclusion": conclusion}
def _conclusions(self, query: str, n: int = 3, classes: list[Classes] = []):
self.dprint("Research", color="green")
with self.dprint.indent_children():
steps = self._query(
self.steps_template(), "1.", r"\d+(?:\.)", query=query, n=n
)
conclusions = [self._conclude_step(s, query, n, classes) for s in steps]
return conclusions
def _collect_data(
self, query: str, klass: Type[Structurer], corpus: list[Document]
):
self.dprint("Collect Data", klass.__name__, color="green")
with self.dprint.indent_children():
data = self.spawn(klass, query=query).extract(corpus)
if data:
return self._query(self.structured_data_template(data), query=query)
else:
return "No relevant data found."
def query(
self,
query: str,
n: int = 3,
classes: list[Classes] = [],
corpus: list[Document] = [],
):
"""Runs the entire question-answering process.
Args:
query: The question to ask.
n: The number of facts to use from the vector store per query.
classes: The interview tags to use as filters (AND).
corpus: The corpus of documents to use for structured data extraction.
Returns:
answer (str): The answer to the question.
"""
answers = [
{
"method": "extract structured data",
"answer": self._collect_data(query, Structurer, corpus),
},
{
"method": "construct a SQL table",
"answer": self._collect_data(query, SQLStructurer, corpus),
},
{
"method": "summarization of facts",
"answer": self._query(
self.conclusions_template(self._conclusions(query, n, classes)),
query=query,
),
},
]
self.dprint(f"Select best answer", query, color="green")
resp = self._query(
self.meta_conclusions_template(answers), query=query, quiet=True
)
if res := re.search(r"```(.*)(```)?", resp, re.DOTALL):
answer = res.group(1).replace("```", "").strip()
elif res := re.search(r"Returned Answer:(.*)(```)?", resp, re.DOTALL):
answer = res.group(1).strip()
else:
answer = resp.replace("```", "").strip()
if (
summary := self.summarizer.summarize_structured_answer(query, answer)
) and summary != answer:
return answer + "\n\n" + summary
else:
return answer
|