Databricks アシスタントとは？

Slide 1

Slide 1 text

Slide 2

Slide 2 text

92% のUSベースの開発者はすでに AIコーディングツールを社内外で活用しています Source: Beet Roots 70% の人はAIコーディングツールは作業での優位性をもたらすと述べており、コードの優れた品質、コンプリーション時間、障害解決について言及しています Source: Github blog 2024

Slide 3

Slide 3 text

企業は自然言語ツールで自分たちのデータ実践者を支援したいと考えています

Slide 4

Slide 4 text

©2024 Databricks Inc. — All rights reserved AIアシスタントは企業データではうまく動作しません適切なデータ資産を特定することができません AIは適切なテーブル、パイプライン、クエリー、ダッシュボードがどこにあるのかを知りません 4 AIはあなた固有のデータ構造、利用パターン、ビジネス用語を理解しません AIはユーザーやグループがどのスキーマ、テーブル、列を参照できるのかを知りません一般的なレスポンスしか生成できませんアクセスポリシーに準拠しません汎用的なLLMはデータインテリジェンスに欠けています

Slide 5

Slide 5 text

©2024 Databricks Inc. — All rights reserved 5 データインテリジェンスはあなたのビジネス固有の文脈を提供しますデータインテリジェンスエンジンはあなたのデータプラットフォームのシグナルを学習するために AIを活用します最近データカタログ文書人気ダッシュボードリネージノートブッククエリー常に適切なデータと AI資産を特定します非常に適切なレスポンスを生成しますユーザーの参照が許可されているデータのみを返却しますポリシーデータインテリジェンスエンジン AIアシスタントはあなた固有のデータとビジネス文脈を考慮します

Slide 6

Slide 6 text

Slide 7

Slide 7 text

Slide 8

Slide 8 text

Slide 9

Slide 9 text

©2024 Databricks Inc. — All rights reserved すべてのユーザー体験でのAIアシスト 9 ダッシュボードでの Databricks アシスタント SQLエディタのDatabricks アシスタントすべてのユーザーの生産性をブーストデータサイエンティスト | データエンジニア | SQLアナリスト | ビジネスユーザーヘルプでの Databricks アシスタントノートブックの Databricks アシスタント

Slide 10

Slide 10 text

Slide 11

Slide 11 text

お客様 ❤ Databricksアシスタント

Slide 12

Slide 12 text

©2024 Databricks Inc. — All rights reserved "At SiriusXM, we're using Databricks Assistant to create notebooks, author complex queries, identify ways to join tables within data lakes, and resolve coding issues, saving our data teams development time." Bernie Graham, VP Data Engineering

Slide 13

Slide 13 text

©2024 Databricks Inc. — All rights reserved A groundbreaking AI-powered tool that transforms coding productivity “ Saptagiri Kintali, Morgan Stanley This AI-based companion is set to reshape the way we code and interact with our lakehouse. “ Jeroen Roosen, Intellus Group You can’t take this away from me! “ Mike Lavina, 84.51 The introduction of the Databricks Assistant has made it easier for our user base to improve their skill set. “ Nicholas Heier, General Motors I was able to code 200+ lines of robust code in a language I've never coded before “ Josue A. Bogran, Kythera Labs “This cutting-edge AI companion has revolutionized my data analysis journey, simplifying complex tasks and accelerating productivity —Byron Exaporriton, ABN AMRO “The convergence of generative AI and data development” —Alaeddin Khader, Core42 “For someone that's competent, it's extremely good at accelerating development” —Luke Woolley, SSE Business Energy

Slide 14

Slide 14 text

Slide 15

Slide 15 text

©2024 Databricks Inc. — All rights reserved 単一のDatabricksアシスタントのUX すべてのDatabricks製品ページで DatabricksアシスタントがAIコンパニオンとして利用できるようになりますジョブの失敗を診断するためにアシスタントを活用し、使用している UI固有のヘルプを参照します文脈を保存するために会話スレッドを作成し、あとで会話を再開します Public Preview

Slide 16

Slide 16 text

Slide 17

Slide 17 text

©2024 Databricks Inc. — All rights reserved Databricksアシスタントの利用レポート組織における生産性の改善や全体的な価値を理解するためにDatabricksアシスタントの利用メトリクスを分析アシスタントのログはアクセスをコントロールするためにUnity Catalogにある既存の system.access.audit テーブルに格納あなたのデータチームで使用されるアシスタント機能や最も共通した利用パターンに対する洞察を提供するための事前構築済みのダッシュボード Public Preview

Slide 18

Slide 18 text

©2024 Databricks Inc. — All rights reserved DBがホストするモデルによるアシスタント Databricksがホスト、管理するモデルによるDatabricksアシスタントの活用 DatabricksIQモデルをサーブするために Databricksモデルサービングを活用。クラウド間のトラフィックを排除します。モデルはステートレスであり、顧客データを記録しません。顧客データはマルチレイヤーの分離によって保護されます。 Private Preview コントロールプレーン Webアプリコンピュートオーケストレーション Unity Catalog クエリーとコード Databricks アシスタント DBIQ モデル

Slide 19

Slide 19 text

©2024 Databricks Inc. — All rights reserved 提案コメントを修正するための文脈追加を追加する能力がメタデータの精度を改善します。 ● コメントに対する質問 ● 別の言語への翻訳 ● 再生成するためのプロンプトコメント位置情報設定のサポートによって、すべてのお客様がこの機能を容易に利用できるようになります。 19 AI生成コメントのエンハンス PuPr *Also available in AWS & Azure GovCloud

Slide 20

Slide 20 text

Slide 21

Slide 21 text

TRY Databricks Assistant TODAY Build your ﬁrst Data+AI project in minutes Take a Tour Watch a Demo Launch Product Tour Launch Demo Video

Slide 22

Slide 22 text

Slide 23

Slide 23 text

Slide 24

Slide 24 text

Enterprises want to understand their business using natural language

Slide 25

Slide 25 text

Slide 26

Slide 26 text

Slide 27

Slide 27 text

©2023 Databricks Inc. — All rights reserved Context is king 👑 “How many DBUs were there in Europe in the last quarter?” • Well-documented Code and Queries: Are there samples we can use as examples? • Popularity: What are the most used tables across my company? • Favorites: Which tables I have used recently? Frequently? • Documentation: Are there relevant wikis and docs for interacting with this data? • Organization: Who accessed these tables? Do I frequently work with them? • Lineage: When was this table last updated? • Dashboards: what are the highly-vetted queries powering frequently-used charts? • UC Descriptions and Tags: What are the tables certiﬁed by my data team?

Slide 28

Slide 28 text

©2023 Databricks Inc. — All rights reserved Semantic Knowledge Graph Data + People + Activity Lineage Lineage Assets Favorites Modiﬁed By Popular Tags Created By Recent View Tables Notebooks Dashboards … Similar Activity User Activity People Terms Org Chart Metadata Data Documentation Jargon

Slide 29

Slide 29 text

Org Chart Unity Catalog Docs Popularity Dashboards Lineage Notebooks Queries Natural Language Queries Search Administration Custom Apps

Slide 30

Slide 30 text

©2023 Databricks Inc. — All rights reserved | This information is confidential, please do not share externally Unity catalog One security and governance model for all data and AI across the organization Unified Governance Cloud Data Lake All Raw Data (Logs, Texts, Audio, Video, Images) Databricks Assistant Code gen, text-to-sql, fix my code Data Asset Discovery Find me the right table Semantic Knowledge Graph (Data, People, Activity) AI Documentation Enrich metadata Gen-AI Platform (LLMs, Vector Index, RAG) Delta Lake One platform to store and manage all structured, semi-structured, and unstructured data DatabricksIQ Personalized Gen-AI Experiences

Slide 31

Slide 31 text

Slide 32

Slide 32 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary Unity Catalog Data Documentation 32 • Auto-generate concise and informative table and column comments for Unity Catalog • Document your backlog of data assets with missing documentation in minutes Public Preview

Slide 33

Slide 33 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary Databricks Assistant 33 • AI-powered Authoring Assistant integrated into the notebook, ﬁle and SQL editors. • Generate, Fix, and Explain code and queries Public Preview

Slide 34

Slide 34 text

Slide 35

Slide 35 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary Help and Support 35 • LLM-powered Help Assistant that provides a conversational interface for documentation and support • Will give accurate answers for Databricks products and technologies such as Spark, Delta Live Tables, DBSQL. • Improved ﬂow for creating support tickets Public Preview

Slide 36

Slide 36 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary Intelligent Search 36 Contextual and accurate search and knowledge card Accurate semantic search using enterprise knowledge graph Summary knowledge card with details to help quickly get to what you are looking for LLM-powered Natural language understanding Public Preview

Slide 37

Slide 37 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary AI Code Suggestions-As-You-Type • “Ghost text” single and multi-line suggestions that automatically appear as you type. • Completions available for Python, Scala, R and SQL. 37 Coming Soon

Slide 38

Slide 38 text

©2023 Databricks Inc. — All rights reserved Centralized AI Governance Unity Catalog provides a secure, governed collaboration layer Available Now • Enable/Disable in account or per-workspace • Assistant only uses table and comment description; it doesn’t look at row-level data • Uses user permissions (e.g., it does not send metadata relating to tables that the user does not have permission to see) H1 2024 • Assistant integrates with audit logs so you can see usage in your workspace • [Potential] Prioritize tables for the Assistant to use. • [Potential] Human-in-the-loop curation/validation

Slide 39

Slide 39 text

©2023 Databricks Inc. — All rights reserved SKU Pricing 39 Data Intelligence Tools $30 USD per user / month [1,2] Boost the productivity of your data and AI teams [1] Trial Period: $0 for first 6 months after GA [2] Active users only (Note: active = generates DBUs in a Workspace - excludes dashboard viewers) [3] AI-generated comments in UC and in-product intelligent search provided at no additional cost [4] Project Genie (aka Data Rooms) not included in DI Tools SKU (future pricing TBD) What’s Included? [3,4] Databricks Assistant for: • Lakeview Dashboards. Generate visuals using natural language. • Notebooks. Create, explain, and fix SQL and Python code using natural language. • SQL Editor. Create, explain, and fix SQL queries using natural language. • Help. Learn, explore, find, troubleshoot, and get support.

Slide 40

Slide 40 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary 40 Roadmap (subject to change) 2023 Databricks Assistant in Public Preview Cell Actions Help Assistant AI-UC-documentation (Preview) Q1 2024 Databricks Assistant (GA) Lakeview dashboard generation (preview) Assistant Audit Logging Intelligent Search (preview) Q2 Project Genie (preview) Ghost Text (preview) AI-UC-documentation (GA) Q3 Expand more cross-product Assistant scenarios Assistant/DatabricksIQ APIs

Slide 41

Slide 41 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary Databricks has opted into the exemption from abuse monitoring and human review program, under which Microsoft does not store any prompts and completions sent to the Azure OpenAI service. Compute Plane The user optionally decides to execute any code Using the Databricks Assistant 41 Assume Role (sts:AssumeRole) Control Plane 3 Azure OpenAI 2 Databricks attaches some metadata to the request and sends it to Azure OpenAI Users Workload 1 (with network isolation) Dedicated Compute Dedicated Compute User selects to diagnose an error or highlights a cell and types a question 1 All trafﬁc between the control plane and Azure OpenAI service is encrypted with TLS 1.2. All data is encrypted at rest. Customers can leverage CMK.

Slide 42

Slide 42 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary • What data is being sent? • Code or queries in the current notebook cell or SQL tab • Table and column names and descriptions • Previous prompt questions • Favourite tables • The “diagnose error” feature also shares the stack trace from the error output • We do not send your query results • Does Azure OpenAI collect my data? • No. Databricks has opted into the exemption from abuse monitoring and human review program, under which Microsoft does not store any prompts and completions sent to the Azure OpenAI service. • Are there any data residency considerations I should be aware of? • We are currently using an Azure OpenAI service deployed in West Europe for all workspaces deployed within a European Geo and one in East US for everyone else. We will continue to evaluate support for other Azure OpenAI regions for future versions to meet latency and data residency requirements. Data FAQs 42

Slide 43

Slide 43 text

©2023 Databricks Inc. — All rights reserved Conﬁdential and Proprietary Model FAQs • What models are you using? • The Databricks Assistant is currently using Azure OpenAI GPT-3.5 as a model. GPT4 is also available to limited preview customers. We’re continually evaluating new models and services (including OpenAI) and may include these in future iterations of the Assistant. • Do you plan to integrate with other models? • Azure OpenAI gave us the fastest path to iteration. However, we’re continually evaluating new models and services (including OpenAI) and may include these in future iterations of the Assistant. • Is my data being used to train models? • No • Will the Databricks Assistant execute dangerous code? • The Assistant will not automatically execute code on your behalf. AI models are error-prone and can make mistakes, misunderstand prompts, hallucinate answers, and introduce bias. You are fully responsible for the code you execute. 43

Slide 44

Slide 44 text

The AI-powered knowledge engine that uniquely understands your business Databricks

Slide 45

Slide 45 text

©2022 Databricks Inc. — All rights reserved DELTA LAKE DLT Uniﬁed Data Intelligence Platform Streaming Events Batch Ingestion and Orchestration Cloud Ingestion Unity Catalog - Data, AI Governance & Lineage Data Intelligence Platform BI/SQL Tools Power BI Delta Ingestion Optimized Spark COPY INTO Auto Loader Internal & External Data Sources Silver Filtered, Cleaned, Augmented Bronze Raw Ingestion and History Business-level Aggregates Gold Spark Structured Stream/Batch Data Science/Machine Learning WorkFlows/Jobs D B S Q L Lakehouse Target Architecture - AWS Data Science Tools Enterprise Data Catalog Optional BI Serving Web Applications Model Serving Auto ML, Gen AI & LLM’s

Slide 46

Slide 46 text

©2023 Databricks Inc. — All rights reserved 46 DatabricksIQ AI-powered data intelligence engine that uniquely understands your business Mosaic AI Delta Live Tables Workﬂows Databricks SQL Unity Catalog Delta Lake Databricks Databricks Data Intelligence Platform AI that understands your data. Intelligence from learning your data and usage patterns. Natural language interfaces for everyone. Democratizes productivity with data and AI for every employee. Predictive optimization of the platform. Automatically tunes and optimizes your workloads and infrastructure.

Slide 47

Slide 47 text

Slide 48

Slide 48 text

©2023 Databricks Inc. — All rights reserved Databricks Assistant Your context-aware AI assistant. Automatically generates SQL and Python, explains complex code, and ﬁxes issues. AI assistance in every user experience; Notebooks, SQL & File editors, Lakeview, Help, and more. Powered by DatabricksIQ to ensure highly relevant responses based on your data and usage.

Slide 49

Slide 49 text

Example: Search with and without DatabricksIQ Without DatabricksIQ With DatabricksIQ