1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
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
|
from __future__ import annotations
import json
import logging
from typing import Literal
import ftfy
import numpy as np
from pydantic import BaseModel
from autogpt.config import Config
from autogpt.core.resource.model_providers import (
ChatMessage,
ChatModelProvider,
EmbeddingModelProvider,
)
from autogpt.processing.text import chunk_content, split_text, summarize_text
from .utils import Embedding, get_embedding
logger = logging.getLogger(__name__)
MemoryDocType = Literal["webpage", "text_file", "code_file", "agent_history"]
class MemoryItem(BaseModel, arbitrary_types_allowed=True):
"""Memory object containing raw content as well as embeddings"""
raw_content: str
summary: str
chunks: list[str]
chunk_summaries: list[str]
e_summary: Embedding
e_chunks: list[Embedding]
metadata: dict
def relevance_for(self, query: str, e_query: Embedding | None = None):
return MemoryItemRelevance.of(self, query, e_query)
def dump(self, calculate_length=False) -> str:
n_chunks = len(self.e_chunks)
return f"""
=============== MemoryItem ===============
Size: {n_chunks} chunks
Metadata: {json.dumps(self.metadata, indent=2)}
---------------- SUMMARY -----------------
{self.summary}
------------------ RAW -------------------
{self.raw_content}
==========================================
"""
def __eq__(self, other: MemoryItem):
return (
self.raw_content == other.raw_content
and self.chunks == other.chunks
and self.chunk_summaries == other.chunk_summaries
# Embeddings can either be list[float] or np.ndarray[float32],
# and for comparison they must be of the same type
and np.array_equal(
self.e_summary
if isinstance(self.e_summary, np.ndarray)
else np.array(self.e_summary, dtype=np.float32),
other.e_summary
if isinstance(other.e_summary, np.ndarray)
else np.array(other.e_summary, dtype=np.float32),
)
and np.array_equal(
self.e_chunks
if isinstance(self.e_chunks[0], np.ndarray)
else [np.array(c, dtype=np.float32) for c in self.e_chunks],
other.e_chunks
if isinstance(other.e_chunks[0], np.ndarray)
else [np.array(c, dtype=np.float32) for c in other.e_chunks],
)
)
class MemoryItemFactory:
def __init__(
self,
llm_provider: ChatModelProvider,
embedding_provider: EmbeddingModelProvider,
):
self.llm_provider = llm_provider
self.embedding_provider = embedding_provider
async def from_text(
self,
text: str,
source_type: MemoryDocType,
config: Config,
metadata: dict = {},
how_to_summarize: str | None = None,
question_for_summary: str | None = None,
):
logger.debug(f"Memorizing text:\n{'-'*32}\n{text}\n{'-'*32}\n")
# Fix encoding, e.g. removing unicode surrogates (see issue #778)
text = ftfy.fix_text(text)
# FIXME: needs ModelProvider
chunks = [
chunk
for chunk, _ in (
split_text(
text=text,
config=config,
max_chunk_length=1000, # arbitrary, but shorter ~= better
tokenizer=self.llm_provider.get_tokenizer(config.fast_llm),
)
if source_type != "code_file"
# TODO: chunk code based on structure/outline
else chunk_content(
content=text,
max_chunk_length=1000,
tokenizer=self.llm_provider.get_tokenizer(config.fast_llm),
)
)
]
logger.debug("Chunks: " + str(chunks))
chunk_summaries = [
summary
for summary, _ in [
await summarize_text(
text=text_chunk,
instruction=how_to_summarize,
question=question_for_summary,
llm_provider=self.llm_provider,
config=config,
)
for text_chunk in chunks
]
]
logger.debug("Chunk summaries: " + str(chunk_summaries))
e_chunks = get_embedding(chunks, config, self.embedding_provider)
summary = (
chunk_summaries[0]
if len(chunks) == 1
else (
await summarize_text(
text="\n\n".join(chunk_summaries),
instruction=how_to_summarize,
question=question_for_summary,
llm_provider=self.llm_provider,
config=config,
)
)[0]
)
logger.debug("Total summary: " + summary)
# TODO: investigate search performance of weighted average vs summary
# e_average = np.average(e_chunks, axis=0, weights=[len(c) for c in chunks])
e_summary = get_embedding(summary, config, self.embedding_provider)
metadata["source_type"] = source_type
return MemoryItem(
raw_content=text,
summary=summary,
chunks=chunks,
chunk_summaries=chunk_summaries,
e_summary=e_summary,
e_chunks=e_chunks,
metadata=metadata,
)
def from_text_file(self, content: str, path: str, config: Config):
return self.from_text(content, "text_file", config, {"location": path})
def from_code_file(self, content: str, path: str):
# TODO: implement tailored code memories
return self.from_text(content, "code_file", {"location": path})
def from_ai_action(self, ai_message: ChatMessage, result_message: ChatMessage):
# The result_message contains either user feedback
# or the result of the command specified in ai_message
if ai_message.role != "assistant":
raise ValueError(f"Invalid role on 'ai_message': {ai_message.role}")
result = (
result_message.content
if result_message.content.startswith("Command")
else "None"
)
user_input = (
result_message.content
if result_message.content.startswith("Human feedback")
else "None"
)
memory_content = (
f"Assistant Reply: {ai_message.content}"
"\n\n"
f"Result: {result}"
"\n\n"
f"Human Feedback: {user_input}"
)
return self.from_text(
text=memory_content,
source_type="agent_history",
how_to_summarize=(
"if possible, also make clear the link between the command in the"
" assistant's response and the command result. "
"Do not mention the human feedback if there is none.",
),
)
def from_webpage(
self, content: str, url: str, config: Config, question: str | None = None
):
return self.from_text(
text=content,
source_type="webpage",
config=config,
metadata={"location": url},
question_for_summary=question,
)
class MemoryItemRelevance(BaseModel):
"""
Class that encapsulates memory relevance search functionality and data.
Instances contain a MemoryItem and its relevance scores for a given query.
"""
memory_item: MemoryItem
for_query: str
summary_relevance_score: float
chunk_relevance_scores: list[float]
@staticmethod
def of(
memory_item: MemoryItem, for_query: str, e_query: Embedding | None = None
) -> MemoryItemRelevance:
e_query = e_query if e_query is not None else get_embedding(for_query)
_, srs, crs = MemoryItemRelevance.calculate_scores(memory_item, e_query)
return MemoryItemRelevance(
for_query=for_query,
memory_item=memory_item,
summary_relevance_score=srs,
chunk_relevance_scores=crs,
)
@staticmethod
def calculate_scores(
memory: MemoryItem, compare_to: Embedding
) -> tuple[float, float, list[float]]:
"""
Calculates similarity between given embedding and all embeddings of the memory
Returns:
float: the aggregate (max) relevance score of the memory
float: the relevance score of the memory summary
list: the relevance scores of the memory chunks
"""
summary_relevance_score = np.dot(memory.e_summary, compare_to)
chunk_relevance_scores = np.dot(memory.e_chunks, compare_to).tolist()
logger.debug(f"Relevance of summary: {summary_relevance_score}")
logger.debug(f"Relevance of chunks: {chunk_relevance_scores}")
relevance_scores = [summary_relevance_score, *chunk_relevance_scores]
logger.debug(f"Relevance scores: {relevance_scores}")
return max(relevance_scores), summary_relevance_score, chunk_relevance_scores
@property
def score(self) -> float:
"""The aggregate relevance score of the memory item for the given query"""
return max([self.summary_relevance_score, *self.chunk_relevance_scores])
@property
def most_relevant_chunk(self) -> tuple[str, float]:
"""The most relevant chunk of the memory item + its score for the given query"""
i_relmax = np.argmax(self.chunk_relevance_scores)
return self.memory_item.chunks[i_relmax], self.chunk_relevance_scores[i_relmax]
def __str__(self):
return (
f"{self.memory_item.summary} ({self.summary_relevance_score}) "
f"{self.chunk_relevance_scores}"
)
|