Oracle AI Vector Search： Getting Started with the Next Big Thing in 23ai [LRN2366].pdf

1、This presentation is intended to outline our general product direction.It is intended for information purposes only,and may not be incorporated into any contract.It is not a commitment to deliver any material,code,or functionality,and should not be reliedupon in making purchasing decisions.The devel

2、opment,release,timing,and pricing of any features or functionality described for Oracles products may change and remains at the sole discretion of Oracle Corporation.Statements in this presentation relating to Oracles future plans,expectations,beliefs,intentions,and prospects are“forward-looking sta

3、tements”and are subject to material risks and uncertainties.A detailed discussion of these factors and other risks that affect our business is contained in Oracles Securities and Exchange Commission(SEC)filings,including our most recent reports on Form 10-K and Form 10-Q under the heading“Risk Facto

4、rs.”These filings are available on the SECs website or on Oracles website at http:/ information in this presentation is current as of September 2024 and Oracle undertakes no duty to update any statement in light of new information or future events.Some regulatory certifications or registrations to p

5、roducts or services referenced herein are held by Cerner Corporation.Cerner Corporation is a wholly owned subsidiary of Oracle.Cerner Corporation is an ONC-certified health IT developer and a registered medical device manufacturer in the United States and other jurisdictions worldwide.The materials

6、in this presentation pertain to Oracle Health,Oracle,Oracle Cerner,and Cerner Enviza which are all wholly owned subsidiaries of Oracle Corporation.Nothing in this presentation should be taken as indicating that any decisions regarding the integration of any EMEA Cerner and/or Enviza entities have be

7、en made where an integration has not already occurred.Oracle CloudWorld Copyright 2024,Oracle and/or its affiliates1Forward-Looking statementsThe following is intended to outline our general product direction.It is intended for information purposes only,and may not be incorporated into any contract.

8、It is not a commitment to deliver any material,code,or functionality,and should not be relied upon in making purchasing decisions.The development,release,timing,and pricing of any features or functionality described for Oracles products may change and remains at the sole discretion of Oracle Corpora

9、tion.The materials in this presentation pertain to Oracle Health,Oracle,Oracle Cerner,and Cerner Enviza which are all wholly owned subsidiaries of Oracle Corporation.Nothing in this presentation should be taken as indicating that any decisions regarding the integration of any EMEA Cerner and/or Envi

10、za entities have been made where an integration has not already occurred.Oracle CloudWorld Copyright 2024,Oracle and/or its affiliates2Safe harbor statementOracle AI Vector SearchGetting Started with the Biggest Thing in 23aiOracle CloudWorld Copyright 2024,Oracle and/or its affiliates3Tirthankar La

11、hiriSenior Vice President,Mission-Critical Data and AI EnginesHeli HelskyahoCEO,Miracle Finland OyCopyright 2024,Oracle and/or its affiliates4Databases have traditionally performed value-based searchesFind revenue by productsCopyright 2024,Oracle and/or its affiliates5There is a growing volume of un

12、structured business data that must be searched by semantics or meaningFind products that match a photo or descriptionThere are many enterprise use-cases for semantic searchCopyright 2024,Oracle and/or its affiliates6Identify defective parts in assembly line before going into productMatch customers t

13、o products for advertisement,recruitment,etc.Biometric pattern recognitionAll of these are scenarios where you need AI-powered search for business dataCopyright 2024,Oracle and/or its affiliates7What is Vector Search?Copyright 2024,Oracle and/or its affiliates8Vector Search is a breakthrough capabil

14、ity for searching data by semantics Vectors are used in AI to capture the semantics of data:Images,documents,videos,or even structured dataA vector is a sequence of numbers,called dimensions,used to capture the important“features”of the dataRepresent the semantic content of data,not the actual words

15、 in a document or pixels in an imageProduced by AI Deep Learning Models known as Embedding ModelsVector3342162150Copyright 2024,Oracle and/or its affiliates9The main operation on vectors is the Mathematical Distance between them82133262Distance(Euclidean Squared)=(3-2)2+(1-6)2+(2-2)2+(8-3)2)There ar

16、e many mathematical distance formulasCopyright 2024,Oracle and/or its affiliates10Laptop Gen 32Support Rep:Jane DoeEmail: Severity 1 Spontaneous reboot Resolved Applied OS Update 42Support IncidentAn Example Business Scenario:The vector for a support incident could be.Vector3342162150Copyright 2024,

17、Oracle and/or its affiliates11Laptop Gen 32Support Rep:Jane DoeEmail: Severity 1 Spontaneous reboot Resolved Applied OS Update 42Support IncidentAn Example Business Scenario:The vector for a support incident could be.Each dimension(number),represents a different feature of the support incidentVector

18、FeaturesSeverityCurrent OS versionSymptomsProductStatus3342162150Copyright 2024,Oracle and/or its affiliates12Note:Features are often chosen by ML algorithms and are not as simple as shown hereSupport incident vectors when collapsed into 2 dimensions instead of hundreds could look like thisd1Support

19、 Incidents for Laptops running slowlySupport Incident for Desktop crashingd2Copyright 2024,Oracle and/or its affiliates13Similarity Property:Support Incidents that are more similar also produce vectors that are closer togetherd1Support Incidents for Laptops running slowlySupport Incident for Desktop

20、 crashingd2Not as SimilarSimilarCopyright 2024,Oracle and/or its affiliates14Arent there dozens of Vector Databases out there already?Copyright 2024,Oracle and/or its affiliates15Why is Oracle adding AI Vector Search?One solution is to continuously send business data to a vector databaseBusiness Dat

21、abaseSearches combining semantic search and value based search on business data requires both types of searches to run togetherCopyright 2024,Oracle and/or its affiliates16Vector Database42162150Incident DataOne solution is to continuously send business data to a vector databaseBusiness DatabaseSear

22、ches combining semantic search and value based search on business data requires both types of searches to run togetherCopyright 2024,Oracle and/or its affiliates17Vector DatabaseIncident Data42162150You need to send a lot of data since you cant predict the question that might be askedCustomer DataPr

23、oduct DataOne solution is to continuously send business data to a vector databaseBusiness DatabaseSearches combining semantic search and value based search on business data requires both types of searches to run togetherCopyright 2024,Oracle and/or its affiliates18Vector DatabaseIncident Data 421621

24、50Causes data staleness,adds complexity,compromises security Customer DataProduct DataOne solution is to continuously send business data to a vector databaseBusiness DatabaseSearches combining semantic search and value based search on business data requires both types of searches to run togetherCopy

25、right 2024,Oracle and/or its affiliates19Vector DatabaseIncident Data42162150Customer DataProduct DataEnterprise DBs typically have an order of magnitude more sophisticated query capabilities,fault-tolerance,security,etc.,than Vector DBsConverged DatabaseMany advantages:All searches run on CURRENT d

26、ata No need to predict what data is neededNo need to move and synchronize dataNo need to manage multiple productsNo compromise to security and fault tolerance Copyright 2024,Oracle and/or its affiliates20Better Approach:Add Vector Search to your Business Database SQL easily supports AI vector search

27、 on unstructured dataCopyright 2024,Oracle and/or its affiliates21Find the top 10 matching incidentsSELECT FROM Support_Incidents ORDER BY VECTOR_DISTANCE(incident_vector,:search_vector)FETCH FIRST 10 ROWS ONLY;Enables combining AI vector search on unstructured data(Incidents)with value-based search

28、 on structured data(Products)Copyright 2024,Oracle and/or its affiliates22Find the top 10 matching incidents for a laptopSELECT FROM Support_Incidents S,Products PWHERE S.prod_id=P.id AND P.type=Laptop ORDER BY VECTOR_DISTANCE(incident_vector,:search_vector)FETCH FIRST 10 ROWS ONLY;Enables combining

29、 AI vector search on unstructured data(Incidents)with value-based search on structured data(Products and Customers)Copyright 2024,Oracle and/or its affiliates23Find the top 10 matching incidents for a laptop reported by customers in Las VegasSELECT FROM Support_Incidents S,Products P,Customers CWHER

30、E S.prod_id=P.id AND S.cust_id=C.id AND P.type=Laptop AND C.city=Las VegasORDER BY VECTOR_DISTANCE(incident_vector,:search_vector)FETCH FIRST 10 ROWS ONLY;Enables combining AI vector search on unstructured Support Datawith value-based search on structured data for Products and CustomersCopyright 202

31、4,Oracle and/or its affiliates24Find the top 10 matching incidents for a laptop reported by customers in Las VegasCombines incident,customer and product data,and AI search in a few lines of SQLA single integrated solution,all data fully consistentAny developer can learn to use it in 5-minutesSELECT

32、FROM Support_Incidents S,Products P,Customers CWHERE S.prod_id=P.id AND S.cust_id=C.id AND P.type=Laptop AND C.city=Las VegasORDER BY VECTOR_DISTANCE(incident_vector,:search_vector)FETCH FIRST 10 ROWS ONLY;Copyright 2024,Oracle and/or its affiliates25What is Oracle AI Vector Search?Encode stored dat

33、a as vectors using your chosen embedding modelEncode search data as a vector using the same modelFind the K nearest stored vectors by distanceReturn the data corresponding to the vectorsAI Vector Search consists of 4 simple steps26Copyright 2024,Oracle and/or its affiliates1234Search Data by Semanti

34、c Similarity rather than by ValuesAI Vector Search HighlightsGenerate vector embeddings from unstructured data Store vectors in table columns using new VECTOR typeBuild approximate vector indexes on VECTOR columnsPerform AI Vector Search on VECTOR columns using SQLScale-Up and Scale-Out to massive d

35、ata and workloads sizesGenerateSearchIndexStoreCopyright 2024,Oracle and/or its affiliates27Scale-OutCopyright 2024,Oracle and/or its affiliates28Generate VectorsAI vector search offers three alternatives for vector embedding generationUse Pre-Created EmbeddingsLoad vectors directly from external fi

36、les into database into VECTOR columns or map the data as external tablesUse an external embedding serviceGenerate embeddings using external callouts via UTL_TO_EMBEDDING()PLSQL functionUse a database resident embedding modelGenerate embeddings using the VECTOR_EMBEDDING()SQL function using an import

37、ed embedding model so that no data leaves the databaseOracle CloudWorld Copyright 2024,Oracle and/or its affiliates29Vector Embedding Generation|Your Way123Generate vectors outside the database with AI model providers like Open-AI,Cohere,and GoogleFirst create a credential using the new CREATE_CREDE

38、NTIAL()API New UTL_TO_EMBEDDING()API to generate vectors outside the databaseOracle CloudWorld Copyright 2024,Oracle and/or its affiliates30/Initialize parameters to specify providervar params CLOB;DBMS_VECTOR.CREATE_CREDENTIAL(“OPENAI_CRED”,auth_params);params:=provider:”openai,credential_name:”OPE

39、NAI_CRED,url:https:/ vectors outside the database with AI model providers like Open-AI,Cohere,and GoogleFirst create a credential using the new CREATE_CREDENTIAL()API Provide that credential along with the provider,to the new UTL_TO_EMBEDDING()APINew UTL_TO_EMBEDDING()API to generate vectors outside

40、 the databaseOracle CloudWorld Copyright 2024,Oracle and/or its affiliates31/Initialize parameters to specify providervar params CLOB;DBMS_VECTOR.CREATE_CREDENTIAL(“OPENAI_CRED”,auth_params);params:=provider:”openai,credential_name:”OPENAI_CRED,url:https:/ vectors from support incidentsSELECT DBMS_V

41、ECTOR_CHAIN.UTL_TO_EMBEDDING(incident_text,json(params)FROM Support_incidents;New VECTOR_EMBEDDING()function to generate vectors within the database/import text model for documentsDBMS_VECTOR.load_onnx_model(model_name=All-MiniLM-L6-v2,model_data=All-MiniLM-L6-v2.onnx .);Copyright 2024,Oracle and/or

42、 its affiliates32Alternately,generate vectors entirely within the databaseOracle Database supports the Open Neural Net Exchange(ONNX)framework to import modelsNew VECTOR_EMBEDDING()function to generate vectors within the databaseCopyright 2024,Oracle and/or its affiliates33Alternately,generate vecto

43、rs entirely within the databaseOracle Database supports the Open Neural Net Exchange(ONNX)framework to import modelsThe VECTOR_EMBEDDING()function can then generate vectors for unstructured data using the imported model/import text model for documentsDBMS_VECTOR.load_onnx_model(model_name=All-MiniLM

44、-L6-v2,model_data=All-MiniLM-L6-v2.onnx .);/generate vectors from support incidentsSELECT VECTOR_EMBEDDING(All-MiniLM-L6-v2 USING incident_text)FROM Support_incidents;Copyright 2024,Oracle and/or its affiliates35StoreVectorsVECTOR Datatype to store and process vectors New VECTOR datatypeFormat for d

45、imension values can be FLOAT32,FLOAT64,and INT8Additional formats coming.36CREATE TABLE Support_Incidents(id number,incident_text CLOB,incident_vector VECTOR(768,FLOAT32);Optional#of dimensionsOptionalformatCopyright 2024,Oracle and/or its affiliatesVECTOR Datatype to store and process vectors New V

46、ECTOR datatypeFormat for dimension values can be FLOAT32,FLOAT64,and INT8Additional formats coming.37CREATE TABLE Support_Incidents(id number,incident_text CLOB,incident_vector VECTOR(768,FLOAT32);Optional#of dimensionsOptionalformatCopyright 2024,Oracle and/or its affiliatesWhy is this needed?Embed

47、ding models are changing constantly but the schema can stay the sameSupport vectors from multiple embedding models in the same column Alternatively,simply specify VECTORCREATE TABLE Support_Incidents(id number,incident_text CLOB,incident_vector VECTOR);Vector Processing Operators The main operation

48、on vectors is VECTOR_DISTANCE(),to find how similar they are Different embedding models can use different distance metrics like EUCLIDEAN,COSINE_SIMILARITY,etc.Regardless of distance metric,all embedding models follow the similarity property:The more similar two items are,the smaller distance betwee

49、n their vectors E.g.VECTOR_DISTANCE(,)VECTOR_DISTANCE(,)VECTOR_DISTANCE(VECTOR1,VECTOR2,)Copyright 2024,Oracle and/or its affiliates38Copyright 2024,Oracle and/or its affiliates40Vector IndexesCopyright 2024,Oracle and/or its affiliates41New vector indexes provide 100 x faster and highly accurate si

50、milarity searches Graph-based index where vertices represent vectors and edges between vertices represent similarityIn-Memory only index-highly efficient for both accuracy and speedVector Indexes Neighbor Graph Vector Index1465327890Graph Vector Index(e.g.HNSW Index)Copyright 2024,Oracle and/or its

51、affiliates42Neighbor Graph Vector IndexIn-memory index designed for speed and accuracyConsidered the“B+tree index for Vectors”Construction The lowest layer of the graph has all the vectors Higher layers have a decaying fraction of vectors Vectors are connected based on similaritySearch Search begins

52、 from the top layer When the nearest vector is found,the search continues in the layer below The search completes in the lowest layer when the Top K nearest vectors to the query vector are foundCopyright 2024,Oracle and/or its affiliates43Multi-layer in-memory graph indexEntry PointLayer 3Layer 2Lay

53、er 1Layer 0Query VectorFind nearestneighbor thengo down a layerFind nearestneighbor thengo down a layerFind nearestneighbor thengo down a layerGreedy search to find nearestK neighborsPartition Vector Index(e.g.IVF_FLAT index)Partition-based index with vectors clustered into table partitions based on

54、 similarityEfficient scale-out index for unlimited data sizeVector Indexes Neighbor Partition Vector IndexCopyright 2024,Oracle and/or its affiliates44Neighbor Partition Vector Index SearchX-axisY-axis#1#2#3#4#5Query Vector1.Group vectors into partitions using OMLs K-means clustering algo(K=5)2.Comp

55、ute distance from query vector to each partitions centroids3.Identify the 2 nearest partitions4.Compute distance from query vector to all points in Cluster#1 and#3 to find Top 5 closest matches(shown in red)2-Dimensional Data-SetCopyright 2024,Oracle and/or its affiliates45Basic index creation synta

56、x:Choosing the ORGANIZATION for an index is simple:If the index data will fit in memory,it is best to use an INMEMORY NEIGHBOR GRAPHElse use NEIGHBOR PARTITIONSThe DISTANCE clause is optional(the default is COSINE)As a rule of thumb,the distance function should be chosen based on the embedding model

57、 used to generate the vectors Vector Index CreationCopyright 2024,Oracle and/or its affiliates47CREATE VECTOR INDEX incident_idx ON support_incidents(incident_vector)ORGANIZATION INMEMORY NEIGHBOR GRAPH|NEIGHBOR PARTITIONSDISTANCE COSINE_SIMILARITY|EUCLIDEAN|HAMMING|.Vector Index Creation Copyright

58、2024,Oracle and/or its affiliates48CREATE VECTOR INDEX incident_idx ON support_incidents(incident_vector)ORGANIZATION INMEMORY NEIGHBOR GRAPH|NEIGHBOR PARTITIONSDISTANCE COSINE_SIMILARITY|EUCLIDEAN|HAMMING.TARGET ACCURACY|ORGANIZATION:If data fits in-memory,use INMEMORY NEIGHBOR GRAPH else use NEIGH

59、BOR PARTITIONSTARGET ACCURACY:Specify the default accuracy(recall)when the index is usedEasiest for users to specify accuracy as a percent instead of index algorithm parametersContinuous calibration used to map target accuracy to low level parameter valuesSpecialists can still specify low-level para

60、meters if they wantCopyright 2024,Oracle and/or its affiliates49Querying VectorsFind the top 5 Customers by similarity with a search photo vector:SELECT id,name,photo FROM Customers ORDER BY VECTOR_DISTANCE(photo_vec,:QUERY_VEC)FETCH FIRST 5 ROWS ONLY Vector Query Copyright 2024,Oracle and/or its af

61、filiates50Vector similarity search queries can easily be combined with relational filters and joins,e.g.Find the top 5 Customers by similarity with a search photo vector who live in San Francisco:SELECT id,name,photo FROM Customers WHERE city=San Francisco ORDER BY VECTOR_DISTANCE(photo_vec,:QUERY_V

62、EC)FETCH APPROXIMATE FIRST 5 ROWS ONLY;Vector Query With Attribute FiltersCopyright 2024,Oracle and/or its affiliates51Vector similarity search queries can easily be combined with relational filters,joins,e.g.Find the top 5 Customers by similarity with a search photo vector who live in San Francisco

63、and who have credit limits greater than$10k based on their status:Vector Query With Attribute Filters and JoinsCopyright 2024,Oracle and/or its affiliates53SELECT id,name,photo FROM Customers c JOIN Status s ON(c.status_id=s.id)WHERE c.city=San Francisco AND s.spending_limit 10000;ORDER BY VECTOR_DI

64、STANCE(c.photo_vec,:QUERY_VEC)FETCH APPROXIMATE FIRST 5 ROWS ONLY;Most enterprise data is normalized,so this is an essential capability Find the top 5 Customers by similarity with a search photo vector:SELECT id,name,photo FROM Customers ORDER BY VECTOR_DISTANCE(photo_vec,:QUERY_VEC)FETCH FIRST 5 RO

65、WS ONLYTARGET ACCURACY|TARGET ACCURACY:Specify the desired accuracy(recall),if different from index accuracySimplicity:Easiest for users to specify accuracy as a percent instead of index search parametersContinuous calibration used to map target accuracy to low-level search parameter valuesSpecialis

66、ts can still specify low-level parameters if they want Vector Query TARGET ACCURACYCopyright 2024,Oracle and/or its affiliates55Find the top 5 Customers by similarity with a search photo vector:SELECT id,name,photo FROM Customers ORDER BY VECTOR_DISTANCE(photo_vec,:QUERY_VEC)FETCH FIRST 5 ROWS ONLYT

67、ARGET ACCURACY|TARGET ACCURACY:Specify the desired accuracy(recall),if different from index accuracySimplicity:Easiest for users to specify accuracy as a percent instead of index search parametersContinuous calibration used to map target accuracy to low-level search parameter valuesSpecialists can s

68、till specify low-level parameters if they want Vector Query TARGET ACCURACYCopyright 2024,Oracle and/or its affiliates56AI Vector Search|TransactionsCopyright 2024,Oracle and/or its affiliates57Oracles AI Vector Search Indexes maintain transactional consistency with DML activity3342162150A new vecto

69、r is insertedNeighbor Partition Vector Index146532870Neighbor Graph Vector Index9Copyright 2024,Oracle and/or its affiliates58Scale-OutAI Vector search transparently scales vector processing across the compute nodes in a RAC clusterWith full data consistencyAI Vector Search|Scale-Out with Real Appli

70、cation Clusters Copyright 2024,Oracle and/or its affiliates59RAC node 1RAC node 2RAC node 3RAC node 1RAC node 2RAC node 3Smart Exadata StorageOracle AI Vector search can be transparently offloaded to smart Exadata storage for faster searchAI Vector Search|Scale-Out with Exadata Smart StorageCopyrigh

71、t 2024,Oracle and/or its affiliates60AI Vector Search|Scale-Out with PartitioningPartition or Sub-Partition by Relational AttributesBuild Vector index on each PartitionFilter by Partition key,perform Vector Search on qualifying partitions onlyCopyright 2024,Oracle and/or its affiliates61Find top 10

72、matching customers in USA Vector index of customer photosAI Vector Search|Scale-Out with ShardingDatabase native sharding enables planet scale vector searchSharding can be used both for unlimited scale or data sovereigntyBuild Vector index on each independent shard database Copyright 2024,Oracle and

73、/or its affiliates62AI Vector Search is fully integrated with Oracle Database 23ai featuresCopyright 2024,Oracle and/or its affiliates63Parallel ExecutionAnalyticsShardingPartitioningSecurityRACTransactionsExadataCopyright 2024,Oracle and/or its affiliates64How fast is Oracle AI Vector Search?Specia

74、list DBs must be faster,right?Oracle CloudWorld Copyright 2024,Oracle and/or its affiliates65Vector Benchmark Performance 8491431877570119489110935558235632795281SIFTGISTFASHION-MISTQPSVECTOR SEARCH PERFORMANCE(QPS 95%RECALL,1 THREAD*)O-DB-VDB-1O-VDB-2O-VDB-3O-DOCDB-4Oracle*Open-Source Databases:O-D

75、B-VDB(Relational DB w/Vectors),O-VDB(Vector DB),O-DOCDB(Document DB)Note:O-DOCDB is 20%faster BUT it uses 2-3X more CPU!Oracle CloudWorld Copyright 2024,Oracle and/or its affiliates66Vector Benchmark Performance 8491431877570119489110935558235632795281SIFTGISTFASHION-MISTQPSVECTOR SEARCH PERFORMANCE

76、(QPS 95%RECALL,1 THREAD*)O-DB-VDB-1O-VDB-2O-VDB-3O-DOCDB-4OracleSIFT1M 128-dim VectorsOracle is 3-6X faster*Open-Source Databases:O-DB-VDB(Relational DB w/Vectors),O-VDB(Vector DB),O-DOCDB(Document DB)Note:O-DOCDB is 20%faster BUT it uses 2-3X more CPU!Oracle CloudWorld Copyright 2024,Oracle and/or

77、its affiliates67Vector Benchmark Performance 8491431877570119489110935558235632795281SIFTGISTFASHION-MISTQPSVECTOR SEARCH PERFORMANCE(QPS 95%RECALL,1 THREAD*)O-DB-VDB-1O-VDB-2O-VDB-3O-DOCDB-4OracleSIFT1M 128-dim VectorsOracle is 3-6X fasterFashion-MNIST60K 768-dim VectorsOracle is 3-10X faster*Open-

78、Source Databases:O-DB-VDB(Relational DB w/Vectors),O-VDB(Vector DB),O-DOCDB(Document DB)Note:O-DOCDB is 20%faster BUT it uses 2-3X more CPU!Oracle CloudWorld Copyright 2024,Oracle and/or its affiliates68Vector Benchmark Performance 8491431877570119489110935558235632795281SIFTGISTFASHION-MISTQPSVECTO

79、R SEARCH PERFORMANCE(QPS 95%RECALL,1 THREAD*)O-DB-VDB-1O-VDB-2O-VDB-3O-DOCDB-4OracleSIFT1M 128-dim VectorsOracle is 3-6X fasterFashion-MNIST60K 768-dim VectorsOracle is 3-10X fasterGIST1M 968-dim VectorsOracle is 2-3X faster than Vector DBs*Open-Source Databases:O-DB-VDB(Relational DB w/Vectors),O-V

80、DB(Vector DB),O-DOCDB(Document DB)Note:O-VDB-3 is 1.27X faster BUT uses 3-4x more CPU Cores!Where to use AI Vector Search?Find Similar Support TicketsBiometric pattern recognitionFind Similar Products Product RecommendationDetect manufacturing anomaliesText search using semantic similarity Enterpris

81、e Similarity Search Use-CasesCopyright 2024,Oracle and/or its affiliates69AI Vector Search is also a critical part of the Generative-AI Ecosystem Copyright 2024,Oracle and/or its affiliates70With the Power of Generative AI,Vector Search allows you to interact with business data using Natural Languag

82、e Copyright 2024,Oracle and/or its affiliates71Copyright 2024,Oracle and/or its affiliates72But why do I need Vector Search?Using Generative AI to answer natural language questionsCopyright 2024,Oracle and/or its affiliates73LLMUserUsing Generative AI to answer natural language questionsCopyright 20

83、24,Oracle and/or its affiliates74Users natural language questions is sent to LLMLLM1Why does my laptop keep rebooting?UserUsing Generative AI to answer natural language questionsCopyright 2024,Oracle and/or its affiliates75LLM12GenAI uses the users question plus its general knowledge to provide an a

84、nswerUsers natural language questions is sent to LLMWhy does my laptop keep rebooting?UserUsing Generative AI to answer natural language questionsA text response is sent back to the user in whatever format was specified123Because it is bad.Maybe buy a new oneWhy does my laptop keep rebooting?UserCop

85、yright 2024,Oracle and/or its affiliates76GenAI uses the users question plus its general knowledge to provide an answerUsers natural language questions is sent to LLMLLMThe LLMs response based on its general knowledge isnt great.Why?Using Generative AI to answer natural language questionsCopyright 2

86、024,Oracle and/or its affiliates77Because it is bad.Maybe buy a new oneWhy does my laptop keep rebooting?LLMUserWhy using Generative AI alone is not good enoughLarge Language Models LLMs are trained on a broad range of data from the internetHowever,they are frozen on a past snapshot of the internet

87、with no access to private enterprise dataLLMs can also hallucinate(i.e.,inaccurate responses)LLMs by themselves therefore often provide poor-quality responses to support questionsCopyright 2024,Oracle and/or its affiliates78Because it is bad.Maybe buy a new oneWhy does my laptop keep rebooting?LLMUs

88、erVector Search improves Generative AI by augmenting prompts with private database content This helps produce better answers to user questionsAvoid having to train LLMs on sensitive enterprise data Called:Retrieval Augmented Generation(RAG)Copyright 2024,Oracle and/or its affiliates7942162550UserA u

89、sers natural language question is encoded as a vector and sent to a Vector Database1Retrieval Augmented Generation works like thisCopyright 2024,Oracle and/or its affiliates80Why does my laptop keep rebooting?33421621542162550Vector DatabaseVector Search finds private content such as incidents that

90、match the users question2UserCopyright 2024,Oracle and/or its affiliates8142162550Vector DatabaseIncident ReportsRetrieval Augmented Generation works like thisThe users question is augmented with previous incident reports3GenAIUserCopyright 2024,Oracle and/or its affiliates8242162550Vector DatabaseI

91、ncident ReportsRetrieval Augmented Generation works like thisGenAI uses the question plus the content and general knowledge to provide an informed answer4UserGenAIRetrieval Augmented Generation works like thisCopyright 2024,Oracle and/or its affiliates8342162550Vector DatabaseIncident ReportsMay nee

92、d to Apply OS update The LLMs response is better thanks to the additional context.But can we do better?UserGenAIRetrieval Augmented Generation works like thisCopyright 2024,Oracle and/or its affiliates8442162550Vector DatabaseIncident ReportsMay need to Apply an OS update Copyright 2024,Oracle and/o

93、r its affiliates85Oracle AI Vector Search takes RAG one step furtherHaving AI vector search in the same database as your Customer and Product data allows a combination of structured business and vector data to be more easily used to augment promptsThis helps produce more accurate answers to user que

94、stionsAlso preserves business security rules:Only data that user is authorized to access,is used in RAG!Copyright 2024,Oracle and/or its affiliates86UserA users natural language question is encoded as a vector and sent to AI Vector Search1RAG with AI Vector Search works like thisCopyright 2024,Oracl

95、e and/or its affiliates87Why does my laptop keep rebooting?334216215AI Vector SearchAI Vector Search finds private content such as incidents for the same product type and customer location that match the users question2UserCopyright 2024,Oracle and/or its affiliates88AI Vector SearchIncident Reports

96、PlusProduct and customer dataRAG with AI Vector Search works like thisThe users question is augmented with the content3GenAIUserCopyright 2024,Oracle and/or its affiliates89AI Vector SearchIncident ReportsPlusProduct and customer dataRAG with AI Vector Search works like thisGenAI uses the question p

97、lus the content and general knowledge to provide a better informed answer4UserGenAIRAG with AI Vector Search works like thisCopyright 2024,Oracle and/or its affiliates90AI Vector SearchThe issue is with the firmware controlling the fan.Apply OS update 42 while plugged in,in a cool air conditioned en

98、vironment to prevent overheatingIncident ReportsPlusProduct and customer dataUserGenAIRAG with AI Vector Search works like thisCopyright 2024,Oracle and/or its affiliates91AI Vector SearchThe LLMs response is much-improved thanks to AI Vector Search and the additional contextThe issue is with the fi

99、rmware controlling the fan.Apply OS update 42 while plugged in,in a cool air conditioned environment to prevent overheatingIncident ReportsPlusProduct and customer dataSome AI Vector Customer Use Cases for Generative AICopyright 2024,Oracle and/or its affiliates92Document ClassificationSupport Chatb

100、otIdentification of PII in documentsCustomer call classificationOnline travel agent chatbotNatural language catalog searchCATALOGDemoChat with APEX using RAGOracle CloudWorld Copyright 2024,Oracle and/or its affiliates93Step 1:Upload PDFs to the DatabaseStep 2:Chat with APEX using GenAI+AI Vector Se

101、archAI Vector Search powers Complete Gen AI pipelineData SourcesDocument LoadersDocument Transformation(e.g.,Text Splitting,Summarization)Embedding ModelsVector DatabaseSimilarity SearchLLMsUserAI Vector Search in Oracle 23ai DatabaseRAG1.Native Database APIs for Gen-AI,from Document Processing to R

102、EST call-outs to LLM Providers2.Tight integration with 3rd party frameworks such as LangchainCopyright 2024,Oracle and/or its affiliates98DBMS_VECTOR and DBMS_VECTOR_CHAINData ManagementVector IndexingIndex Creation(IVF/HNSW)Index RebuildSimilarity SearchSimplified Top K query API just like niche ve

103、ctor databasesSpecify Query Vector,K,select list columns,accuracy,etc.ONNX Model ManagementLoad/Drop ONNX modelWraps DBMS_DATA_MINING proceduresAccuracy ReportingTwo use-cases:Measure accuracy of a vector index searchTweak query parameters to achieve desired accuracyIndex Accuracy ReportAccuracy ach

104、ieved for auto-captured query vectorsIndex Query AccuracyAccuracy achieved given a Target Accuracy%,or low-level index search parametersChainable FunctionsEnables end-to-end data transformation and search pipelines(like Langchain)Creating Chunks from Text documentsEmbedding Generation using ONNX mod

105、els or external service providers(e.g.,Cohere,Open AI,Gemini)using RESTSummarization using Oracle Text“Gist”or external summarization modelsPL/SQL packages for managing vectors and interacting with Gen AI ecosystemCopyright 2024,Oracle and/or its affiliates99Copyright 2024,Oracle and/or its affiliat

106、es100Key TakeawaysNEW Vector data type for storing vectorsNEW SQL syntax and functions express similarity search with easeNEW vector indexes for high-performanceAI Vector Search is FAST!AI Vector Search powers Complete Gen AI pipelineOracle brings AI vector search to your business data accelerating

107、the development of AI-enabled applicationsCopyright 2024,Oracle and/or its affiliates101AI Vector Search Sessions at OCW 2024Oracle CloudWorld Copyright 2024,Oracle and/or its affiliates102SessionSpeakersDate/TimeLocationChoosing the Right AI Capabilities in Oracle Database 23ai for Your Use Case LR

108、N3350Maria ColganShasank ChavanMark HornickTuesday,Sep 10 11:45 AM-12:30 PM PDTGalileo 902,The Venetian,Level 1AI Vector Search:We Start Simple,Then Dive in Deep LRN2956Shasank ChavanAurosish MishraTuesday,Sep 10 2:45 PM-3:30 PM PDTBallroom G,The Venetian Level 2AI Vector Search:A RAGs to Riches Sto

109、ry LRN3006Aurosish MishraAlexandra CzarlinskaWednesday,Sep 11 11:00 AM-11:45 AM PDTGalileo 1007,The Venetian,Level 1How NRI uses Oracle 23ai AI Vector search for carbon-information tracing LRN1083Shasank ChavanShinichiro OtsukaWednesday,Sep 11 1:00 PM-1:45 PM PDTMarco Polo 706,The Venetian,Level 1Or

110、acle AI Vector Search:Getting Started with the Next Big Thing in 23ai LRN2366Tirthankar LahiriHeli HalskyahoWednesday,Sep 11 3:30 PM-4:15 PM PDTMarco Polo 807,The Venetian,Level 1Oracle and NVIDIA Team Up to Accelerate AI Vector Search LRN3902Shasank ChavanNathan StephensWeiwei GongWednesday,Sep 114

111、:45 PM-5:30 PM PDTBellini 2103,The Venetian,Level 2 AI Vector Search:Powering the Modern Enterprise LRN1191Aurosish MishraRishin MitraThursday,Sep 12 10:15 AM-11:00 AM PDTGalileo 906,The Venetian,Level 1AI Vector Search Training SessionsOracle CloudWorld Copyright 2024,Oracle and/or its affiliates10

112、3SessionSpeakersDate/TimeLocationOracle AI Vector Search:From Zero to Vector in Three Steps THR2976Doug HoodTuesday,Sep 1011:10 AM-11:30 AM PDTCW Hub,Technology Theater 1Getting Started with Oracle AI Vector Search-Offering 1 HOL2974Doug HoodRanjan PriyadarshiTuesday,Sep 102:45 PM-4:15 PM PDTMarco P

113、olo 703,The Venetian,Level 1Building Generative AI Applications Using LangChain with Oracle AI Vector Search HOL2975Doug HoodRanjan PriyadarshiWednesday,Sep 1110:30 AM-12:00 PM PDTCasanova 607,The Venetian,Level 1Building Enterprise RAG with AI Vector Search HOL2973Doug HoodRanjan PriyadarshiWednesd

114、ay,Sep 113:00 PM-4:30 PM PDTCasanova 606,The Venetian,Level 1Getting Started with Oracle AI Vector SearchOffering 2 HOL3603Doug HoodRanjan PriyadarshiThursday,Sep 129:00 AM-10:30 AM PDTCasanova 601,The Venetian,Level 1AI Vector Search|How to get startedCopyright 2024,Oracle and/or its affiliates104V

115、isit https:/ 2024,Oracle and/or its affiliates105Try AI Vector SearchLive Labs https:/ Vector Search|LinksVisit home page:https:/ Vector Search LiveLabs:https:/ Vector Search User Guide:https:/ Vector Search Blog:https:/ RAG bit.ly/oracleai_demoAlways Free Autonomous Database:https:/ Database 23ai Container Image:https:/ 23ai Free:https:/ 2024,Oracle and/or its affiliates106Thank YouOracle CloudWorld Copyright 2024,Oracle and/or its affiliates107