1. \u6570\u636e\u4e0a\u4f20<\/strong><\/p>\n\n\n\n\u5728\u6211\u4eec\u7684\u6848\u4f8b\u4e2d\uff0c\u5e73\u53f0\u53ef\u4ee5\u5c06\u89c6\u9891\u6587\u4ef6\uff08mp4\/mov \u683c\u5f0f\uff09\u4e0a\u4f20\u81f3 Amazon S3 \u8def\u5f84\u4e2d\uff0c\u4f8b\u5982 s3:\/\/<bucket-name>\/video-input\/\uff0c\u4f5c\u4e3a\u540e\u7eed\u5904\u7406\u7684\u6e90\u6570\u636e\u3002<\/p>\n\n\n\n
2. \u89c6\u9891\u5904\u7406\u4e0e\u89c6\u9891 Json \u6587\u4ef6\u751f\u6210<\/strong><\/p>\n\n\n\n\u5f53 Amazon S3 \u8def\u5f84\u4e2d\u6709\u6587\u4ef6\u4e0a\u4f20\u5b8c\u6bd5\u65f6\uff0cAmazon EventBridge \u4f1a\u76d1\u6d4b\u5230\u76f8\u5173\u4e8b\u4ef6\u5e76\u89e6\u53d1 Amazon Lambda \u8c03\u7528 Amazon Bedrock API \u5bf9\u89c6\u9891\u6587\u4ef6\u6309\u7167\u5b9e\u73b0\u521b\u5efa\u7684 DBA \u9879\u76ee\u8fdb\u884c\u5904\u7406\u3002DBA \u9879\u76ee\u662f\u5229\u7528 Lambda \u5728\u89c6\u9891\u5904\u7406\u6d41\u7a0b\u524d\u521b\u5efa\u3002\u89c6\u9891\u5904\u7406\u5b8c\u6210\u540e\uff0c\u751f\u6210\u76f8\u5e94\u7684 json \u6587\u4ef6\uff0c\u5e76\u5c06\u5176\u4f20\u56de Amazon S3 \u5b58\u50a8 s3:\/\/<bucket-name>\/video-output\/\u3002\u76f8\u5173\u793a\u4f8b\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n\n
# invokes bda by async approach with a given input file<\/code>def invoke_insight_generation_async(<\/code> input_s3_uri,<\/code> output_s3_uri,<\/code> data_project_arn, blueprints = None):<\/code> # Extract account ID from project ARN<\/code> account_id = data_project_arn.split(':')[4]<\/code> <\/code> # Construct default Profile ARN using fixed us-west-2 region\uff0cwhich currently supports the 2 models required in the following process<\/code> profile_arn = f\"arn:aws:bedrock:us-west-2:{account_id}:data-automation-profile\/us.data-automation-v1\"<\/code> payload = {<\/code> \"inputConfiguration\": {<\/code> \"s3Uri\": input_s3_uri<\/code> },<\/code> \"outputConfiguration\": {<\/code> \"s3Uri\": output_s3_uri<\/code> },<\/code> \"dataAutomationConfiguration\": {<\/code> \"dataAutomationProjectArn\": data_project_arn,<\/code> \"stage\": \"LIVE\"<\/code> },<\/code> \"dataAutomationProfileArn\": profile_arn,<\/code> \"notificationConfiguration\": {<\/code> \"eventBridgeConfiguration\": {\"eventBridgeEnabled\": True},<\/code> }<\/code> }<\/code> try:<\/code> response = bda_client_runtime.invoke_data_automation_async(**payload)<\/code> print(f\"Successfully invoked data automation: {response}\")<\/code> return response<\/code> except Exception as e:<\/code> print(f\"Error invoking data automation: {str(e)}\")<\/code> raise<\/code><\/pre>\n\n\n\n3. \u57fa\u4e8e\u89c6\u9891 Json \u6587\u4ef6\u7684 Embedding \u5904\u7406<\/strong><\/p>\n\n\n\n\u4e4b\u540e\uff0cAmazon S3 \u6587\u4ef6\u751f\u6210\u7684\u89c4\u5219\u4f1a\u518d\u6b21\u89e6\u53d1\u7684 Amazon EventBridge\uff0c\u4f7f\u5176\u542f\u52a8 Amazon Lambda \u8c03\u7528 Amazon Bedrock \u96c6\u6210\u7684 amazon.titan-embed-text-v2:0 \u6a21\u578b\uff0c\u4e3a Json \u6587\u4ef6\u5185\u5bb9\u505a \u5411\u91cf\u5316\u7528\u4e8e\u540e\u7eed\u67e5\u8be2\u3002<\/p>\n\n\n\n
try:<\/code> response = bedrock_client.invoke_model(<\/code> modelId=\"amazon.titan-embed-text-v2:0\",<\/code> contentType=\"application\/json\",<\/code> accept=\"application\/json\",<\/code> body=json.dumps({<\/code> \"inputText\": text<\/code> })<\/code> )<\/code><\/pre>\n\n\n\n\u7531\u4e8e\u6bcf\u4e2a chapter \u5185\u7684 transcripts \u8f83\u957f\uff0c\u4e3a\u8fbe\u5230\u66f4\u597d\u7684\u68c0\u7d22\u6548\u679c\uff0c\u4f1a\u5bf9\u5176\u8fdb\u884c\u5207\u7247\u5904\u7406\uff0c\u6700\u540e\u5c06\u6574\u7406\u597d\u7684\u6570\u636e\u7ed3\u6784\u5b58\u5165 DocumentDB\u3002\u7531\u4e8e\u662f\u7b80\u6613\u73af\u5883\uff0c\u6211\u4eec\u57fa\u4e8e\u5b57\u7b26\u957f\u5ea6\u505a\u4e86\u7b80\u5355\u7684\u5207\u7247\u5904\u7406\uff1a<\/p>\n\n\n\n
def split_transcript_into_chunks(self, transcript_text, max_chunk_size=500, min_chunk_size=100):<\/code> if not transcript_text:<\/code> return []<\/code> # \u5982\u679c\u8f93\u5165\u662f\u5b57\u5178\uff0c\u5c1d\u8bd5\u63d0\u53d6\u6587\u672c<\/code> if isinstance(transcript_text, dict):<\/code> if 'text' in transcript_text:<\/code> transcript_text = transcript_text['text']<\/code> elif 'representation' in transcript_text and 'text' in transcript_text['representation']:<\/code> transcript_text = transcript_text['representation']['text']<\/code> else:<\/code> transcript_text = json.dumps(transcript_text)<\/code> # \u786e\u4fdd\u8f93\u5165\u662f\u5b57\u7b26\u4e32<\/code> if not isinstance(transcript_text, str):<\/code> transcript_text = str(transcript_text)<\/code> # \u6309\u53e5\u5b50\u5206\u5272\uff08\u53e5\u53f7\u3001\u95ee\u53f7\u3001\u611f\u53f9\u53f7\u540e\u8ddf\u7a7a\u683c\uff09<\/code> # \u4f7f\u7528\u66f4\u5bbd\u677e\u7684\u6a21\u5f0f\uff0c\u5141\u8bb8\u53e5\u5b50\u7ed3\u675f\u540e\u6709\u4efb\u4f55\u6570\u91cf\u7684\u7a7a\u683c<\/code> sentences = re.split(r'([.!?])\\s*', transcript_text)<\/code> # \u5904\u7406\u5206\u5272\u7ed3\u679c\uff0c\u5c06\u6807\u70b9\u7b26\u53f7\u91cd\u65b0\u9644\u52a0\u5230\u53e5\u5b50<\/code> processed_sentences = []<\/code> i = 0<\/code> while i < len(sentences):<\/code> sentence = sentences[i]<\/code> # \u5982\u679c\u662f\u6807\u70b9\u7b26\u53f7\uff0c\u9644\u52a0\u5230\u524d\u4e00\u4e2a\u53e5\u5b50<\/code> if i > 0 and sentence in ['.', '!', '?']:<\/code> processed_sentences[-1] += sentence<\/code> else:<\/code> processed_sentences.append(sentence)<\/code> i += 1<\/code> # \u5c06\u53e5\u5b50\u7ec4\u5408\u6210\u5757<\/code> chunks = []<\/code> current_chunk = \"\"<\/code> for sentence in processed_sentences:<\/code> # \u8df3\u8fc7\u7a7a\u53e5\u5b50<\/code> if not sentence.strip():<\/code> continue<\/code> # \u5982\u679c\u5f53\u524d\u5757\u52a0\u4e0a\u65b0\u53e5\u5b50\u4f1a\u8d85\u8fc7\u6700\u5927\u5927\u5c0f\uff0c\u4fdd\u5b58\u5f53\u524d\u5757<\/code> if current_chunk and len(current_chunk) + len(sentence) + 1 > max_chunk_size:<\/code> if len(current_chunk) >= min_chunk_size:<\/code> chunks.append(current_chunk.strip())<\/code> current_chunk = sentence<\/code> else:<\/code> if current_chunk:<\/code> current_chunk += \" \" + sentence<\/code> else:<\/code> current_chunk = sentence<\/code> # \u6dfb\u52a0\u6700\u540e\u4e00\u4e2a\u5757<\/code> if current_chunk and len(current_chunk) >= min_chunk_size:<\/code> chunks.append(current_chunk.strip())<\/code> return chunks<\/code><\/pre>\n\n\n\n\u5c06 Json \u6587\u4ef6\u5411\u91cf\u5316\u540e\uff0c\u6211\u4eec\u5c06 Json \u6587\u4ef6\u7684\u6570\u636e\u683c\u5f0f\u8fdb\u884c\u4e86\u5904\u7406\uff0c\u4f7f\u5176\u5728 Amazon DocumentDB \u4e2d\u80fd\u66f4\u597d\u5730\u505a\u68c0\u7d22\u3002<\/p>\n\n\n\ndef flatten_video_data(self, video_data, video_name):<\/code> flattened_data = []<\/code> <\/code> # \u5904\u7406\u7ae0\u8282<\/code> chapters = video_data.get('chapters', [])<\/code> for chapter in chapters:<\/code> chapter_summary = chapter.get('chapter_summary', {})<\/code> flattened_chapter_summary = {<\/code> \"video_name\": video_name,<\/code> \"source\": f\"chapter_{chapter.get('chapter_index', 0)}_summary\",<\/code> \"text\": chapter_summary.get('text', \"\"),<\/code> \"embedding\": chapter_summary.get('embedding', []),<\/code> \"start_timestamp_millis\": chapter.get('start_timestamp_millis'),<\/code> \"end_timestamp_millis\": chapter.get('end_timestamp_millis')<\/code> }<\/code> flattened_data.append(flattened_chapter_summary)<\/code> # \u5904\u7406\u7ae0\u8282\u8f6c\u5f55\u5757<\/code> transcript_chunks = chapter.get('transcript_chunks', [])<\/code> for chunk in transcript_chunks:<\/code> flattened_chunk = {<\/code> \"video_name\": video_name,<\/code> \"source\": f\"chapter_{chapter.get('chapter_index', 0)}_transcript_chunk_{chunk.get('chunk_index', 0)}\",<\/code> \"text\": chunk.get('text', \"\"),<\/code> \"embedding\": chunk.get('embedding', []),<\/code> \"start_timestamp_millis\": chapter.get('start_timestamp_millis'),<\/code> \"end_timestamp_millis\": chapter.get('end_timestamp_millis')<\/code> }<\/code> flattened_data.append(flattened_chunk)<\/code> return flattened_data<\/code><\/pre>\n\n\n\n\u5bf9\u89c6\u9891\u751f\u6210\u7684\u6587\u5b57\u505a\u5b8c\u5411\u91cf\u5316\u548c\u6570\u636e\u683c\u5f0f\u7684\u6574\u7406\u540e\uff0c\u4fbf\u53ef\u4ee5\u5c06\u5176\u5b58\u5165 Amazon DocumentDB \u4e3a\u7528\u6237\u7684\u68c0\u7d22\u505a\u51c6\u5907\u4e86\u3002<\/p>\n\n\n\n
4. \u7528\u6237\u641c\u7d22\u6587\u5b57\u7684 Embedding \u8f6c\u6362<\/strong><\/p>\n\n\n\n\u5c06\u5411\u91cf\u548c\u6587\u5b57\u5b58\u5165 Amazon DocumentDB \u540e\uff0c\u7528\u6237\u4fbf\u53ef\u4ee5\u4ece\u524d\u7aef\u8f93\u5165\u641c\u7d22\u6587\u5b57\u5e76\u9009\u62e9\u641c\u7d22\u6a21\u5f0f\uff08\u6587\u672c\u641c\u7d22\u6216\u573a\u666f\u641c\u7d22\uff09\uff0c\u7cfb\u7edf\u4f1a\u4f7f\u7528 amazon.titan \u2013 embed \u2013 text \u2013 v2:0 \u6a21\u578b\u5bf9\u641c\u7d22\u6587\u5b57\u8fdb\u884c\u5411\u91cf\u5316\uff0c\u4ee5\u4fbf\u540e\u7eed\u8fdb\u884c\u5411\u91cf\u5339\u914d\u3002<\/p>\n\n\n\n
flowchart LR<\/code> A([\u63a5\u6536\u641c\u7d22\u8bf7\u6c42])-->B[\u89e3\u6790\u67e5\u8be2\u53c2\u6570]<\/code> B-->C{\u6267\u884c\u6df7\u5408\u641c\u7d22}<\/code> C-->D[\u5411\u91cf\u641c\u7d22]<\/code> C-->E[\u6587\u672c\u641c\u7d22]<\/code> D-->F([\u5408\u5e76\u641c\u7d22\u7ed3\u679c])<\/code> E-->F<\/code> F-->G([\u4f7f\u7528Cohere\u91cd\u6392\u5e8f])<\/code> G-->H([\u8fd4\u56de\u6700\u7ec8\u7ed3\u679c])<\/code><\/pre>\n\n\n\n\u53cc\u8def\u53ec\u56de\u4e0e\u7ed3\u679c\u91cd\u6392\u5e8f<\/strong>\uff1aAmazon Lambda \u5229\u7528\u8f6c\u6362\u540e\u7684\u5411\u91cf\uff0c\u540c\u65f6\u8fdb\u884c\u6587\u5b57\u548c\u5411\u91cf\u7684\u53cc\u8def\u53ec\u56de\u3002\u5728\u5411\u91cf\u68c0\u7d22\u4e2d\uff0c\u4f7f\u7528 Amazon Document DB \u7684\u805a\u5408 pipeline \u5bf9\u5411\u91cf\u76f8\u4f3c\u5ea6\u8fdb\u884c\u641c\u7d22\uff0c\u6839\u636e\u641c\u7d22\u6a21\u5f0f\u8fc7\u6ee4\u7ed3\u679c\uff0c\u6700\u540e\u901a\u8fc7 project \u8fd4\u56de\u76f8\u5e94\u5b57\u6bb5\uff1b<\/p>\n\n\n\n$project\uff1a\u8fd4\u56de\u76f8\u5e94\u5b57\u6bb5\u5982\u4e0b\uff1a<\/p>\n\n\n\n
text: \"<sample_text>\"<\/code>video_name: \"<sample_name>.mov\" <\/code>source: \"chapter_6_summary\"<\/code>start_timestamp_millis: 181666<\/code>end_timestamp_millis: 227100<\/code><\/pre>\n\n\n\nsimilarity\uff1a\u4e0e\u5411\u91cf\u7d22\u5f15\u4e00\u6837\uff0c\u6211\u4eec\u9009\u62e9\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff08cosine\uff09\u6765\u8861\u91cf\u67e5\u8be2\u5411\u91cf\u548c\u6587\u6863\u5411\u91cf\u4e4b\u95f4\u7684\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n
top_k\uff1a\u8fd4\u56de\u7684\u89c6\u9891\u6570<\/p>\n\n\n\n
filter_condition\uff1a\u7531\u4e8e\u6211\u4eec\u7684\u641c\u7d22\u5206\u4e3a\u573a\u666f\u641c\u7d22\u548c\u53f0\u8bcd\u6587\u672c\u641c\u7d22\uff0c\u4fbf\u4ee5\u6b64\u4e3a\u7b5b\u9009\uff0c\u4ece\u800c\u7528\u6237\u53ef\u4ee5\u81ea\u884c\u9009\u62e9\u641c\u7d22\u573a\u666f\u3002<\/p>\n\n\n\n
pipeline = [<\/code> {<\/code> \"$search\": {<\/code> \"vectorSearch\": {<\/code> \"vector\": query_embedding,<\/code> \"path\": \"embedding\",<\/code> \"similarity\": \"cosine\",<\/code> \"k\": top_k * 3, # Fetch more results since we'll filter them<\/code> \"efSearch\": 64<\/code> }<\/code> }<\/code> },<\/code> # Add a $match stage to filter results based on search mode (scene\/transcripts)<\/code> {<\/code> \"$match\": filter_condition<\/code> },<\/code> # Limit to top_k results after filtering<\/code> {<\/code> \"$limit\": top_k<\/code> },<\/code> {<\/code> \"$project\": {<\/code> \"text\": 1,<\/code>
![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/e7\/e7a8f1471bab87358413b87e12bbb856.png\")
<\/figure>\n\n\n\n![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/90\/9024d8f48c4a4bf409d4c54c4876fd80.png\")
<\/figure>\n\n\n\n![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/35\/35a8e3ad8c6eeebecf75ae7f2dba07a2.png\")
<\/figure>\n\n\n\n![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/63\/63a0527ac57b966b2c937fd01826be84.png\")
<\/figure>\n\n\n\n![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/5a\/5a7c9ccd0d49b649fa468829f5104e0d.png\")
<\/figure>\n\n\n\n![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/ff\/ffc7a339fbf6d881019d4a383e8a29f3.png\")
<\/figure>\n\n\n\n![\"\"\/](\"https:\/\/static001.geekbang.org\/infoq\/ea\/ea5d9e6e387130466b4d86ec357648cf.png\")
<\/figure>\n\n\n\n