-
Notifications
You must be signed in to change notification settings - Fork 5k
/
Copy pathtranscribe_media.ts
225 lines (202 loc) · 6.31 KB
/
transcribe_media.ts
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
import { composeContext } from "@elizaos/core";
import { generateText } from "@elizaos/core";
import { parseJSONObjectFromText } from "@elizaos/core";
import {
Action,
ActionExample,
Content,
HandlerCallback,
IAgentRuntime,
Memory,
ModelClass,
State,
} from "@elizaos/core";
export const transcriptionTemplate = `# Transcription of media file
{{mediaTranscript}}
# Instructions: Return only the full transcript of the media file without any additional context or commentary.`;
export const mediaAttachmentIdTemplate = `# Messages we are transcribing
{{recentMessages}}
# Instructions: {{senderName}} is requesting a transcription of a specific media file (audio or video). Your goal is to determine the ID of the attachment they want transcribed.
The "attachmentId" is the ID of the media file attachment that the user wants transcribed. If not specified, return null.
Your response must be formatted as a JSON block with this structure:
\`\`\`json
{
"attachmentId": "<Attachment ID>"
}
\`\`\`
`;
const getMediaAttachmentId = async (
runtime: IAgentRuntime,
message: Memory,
state: State
): Promise<string | null> => {
state = (await runtime.composeState(message)) as State;
const context = composeContext({
state,
template: mediaAttachmentIdTemplate,
});
for (let i = 0; i < 5; i++) {
const response = await generateText({
runtime,
context,
modelClass: ModelClass.SMALL,
});
console.log("response", response);
const parsedResponse = parseJSONObjectFromText(response) as {
attachmentId: string;
} | null;
if (parsedResponse?.attachmentId) {
return parsedResponse.attachmentId;
}
}
return null;
};
const transcribeMediaAction = {
name: "TRANSCRIBE_MEDIA",
similes: [
"TRANSCRIBE_AUDIO",
"TRANSCRIBE_VIDEO",
"MEDIA_TRANSCRIPT",
"VIDEO_TRANSCRIPT",
"AUDIO_TRANSCRIPT",
],
description:
"Transcribe the full text of an audio or video file that the user has attached.",
validate: async (
_runtime: IAgentRuntime,
message: Memory,
_state: State
) => {
if (message.content.source !== "discord") {
return false;
}
const keywords: string[] = [
"transcribe",
"transcript",
"audio",
"video",
"media",
"youtube",
"meeting",
"recording",
"podcast",
"call",
"conference",
"interview",
"speech",
"lecture",
"presentation",
];
return keywords.some((keyword) =>
message.content.text.toLowerCase().includes(keyword.toLowerCase())
);
},
handler: async (
runtime: IAgentRuntime,
message: Memory,
state: State,
options: any,
callback: HandlerCallback
) => {
state = (await runtime.composeState(message)) as State;
const callbackData: Content = {
text: "", // fill in later
action: "TRANSCRIBE_MEDIA_RESPONSE",
source: message.content.source,
attachments: [],
};
const attachmentId = await getMediaAttachmentId(
runtime,
message,
state
);
if (!attachmentId) {
console.error("Couldn't get media attachment ID from message");
return;
}
const attachment = state.recentMessagesData
.filter(
(msg) =>
msg.content.attachments &&
msg.content.attachments.length > 0
)
.flatMap((msg) => msg.content.attachments)
.find(
(attachment) =>
attachment.id.toLowerCase() === attachmentId.toLowerCase()
);
if (!attachment) {
console.error(`Couldn't find attachment with ID ${attachmentId}`);
return;
}
const mediaTranscript = attachment.text;
callbackData.text = mediaTranscript.trim();
// if callbackData.text is < 4 lines or < 100 words, then we we callback with normal message wrapped in markdown block
if (
callbackData.text &&
(callbackData.text?.split("\n").length < 4 ||
callbackData.text?.split(" ").length < 100)
) {
callbackData.text = `Here is the transcript:
\`\`\`md
${mediaTranscript.trim()}
\`\`\`
`;
await callback(callbackData);
}
// if text is big, let's send as an attachment
else if (callbackData.text) {
const transcriptFilename = `content/transcript_${Date.now()}`;
// save the transcript to a file
await runtime.cacheManager.set(
transcriptFilename,
callbackData.text
);
await callback(
{
...callbackData,
text: `I've attached the transcript as a text file.`,
},
[transcriptFilename]
);
} else {
console.warn(
"Empty response from transcribe media action, skipping"
);
}
return callbackData;
},
examples: [
[
{
user: "{{user1}}",
content: {
text: "Please transcribe the audio file I just sent.",
},
},
{
user: "{{user2}}",
content: {
text: "Sure, I'll transcribe the full audio for you.",
action: "TRANSCRIBE_MEDIA",
},
},
],
[
{
user: "{{user1}}",
content: {
text: "Can I get a transcript of that video recording?",
},
},
{
user: "{{user2}}",
content: {
text: "Absolutely, give me a moment to generate the full transcript of the video.",
action: "TRANSCRIBE_MEDIA",
},
},
],
] as ActionExample[][],
} as Action;
export default transcribeMediaAction;