Turning selected company documents into business definitions for data-model metadata and conversational analytics.Illustrative AI-generated image: selected business documents become grouped contextual definitions, with unresolved items kept aside for review.Can't get the business to spend time…
Annons
Annons
Turning selected company documents into business definitions for data-model metadata and conversational analytics.Illustrative AI-generated image: selected business documents become grouped contextual definitions, with unresolved items kept aside for review.Can't get the business to spend time building a glossary for you? You explain the benefits to them, provide the templates, give examples, have meetings, set deadlines, and...still nothing.I made a workflow in Google Gemini Agent Flow Builder to mine selected company documents for business context. It took about an hour to build and another hour to test, and supplied hundreds of well-defined business entities to augment our existing context. We use that context heavily in data-model metadata generation and conversational analytics.Google's current docs call this Workflow Builder in Gemini Enterprise; Gemini Agent Flow Builder is the name I used in September. What it is called when you read this article, I could not say, but that is why we love Google.Once your administrator has enabled Workflow Builder, open the Gemini Enterprise app and go to Agents > Create agent. Describe what you want in plain language, then inspect and refine the generated workflow and test it against your documents.Choose the documents that are relevant to your contextI used selected folders in Google Drive. Start with the question the data model or conversational system needs to answer, then choose the folders and subdirectories that contain relevant documents.For our technology team’s chatbot, that can mean project business cases, design and architecture documents, project records, business-unit process documents, and meeting minutes. Company communications and relevant company website content can help when they use the same business language outside the project folders. Marketing material probably won’t add much unless it explains something the chatbot genuinely needs to answer.Set up the workflow knowledge with the files or folders for your context. Include the relevant subdirectories and check that their documents appear in the run’s results.Enabling the full Drive tool on the same agent step also overrides the restricted selection, so keep that setting in view when setting up the source scope. Gemini Agent Flow Builder’s knowledge settings support selected Drive and OneDrive resources.OneDrive and SharePoint are also connector options. For documents in Google Cloud Storage or Windows folders, check how they can be made available through a supported connection; neither is a direct connector in the Gemini Agent Flow Builder list.On AWS, Bedrock Knowledge Bases can synchronise an S3 prefix. That will be a separate AWS setup, however, with its own access requirements.Describe the workflow you wantAsk the builder to include inputs for the root folder and output destination: a folder for a new glossary file, or an existing glossary file to extend.Describe the processing as a sequence: extract terms and meanings from the selected documents, compare repeated terms in their business context, consolidate matching definitions, merge the existing glossary if supplied, then check consistency and record issues. Gemini Agent steps can pass structured records between steps, so extraction feeds consolidation and the glossary merge.Give the extraction step a template so it returns the same fields for each definition. That makes it easier to compare meanings and combine the outputs.Illustrative output record (fictional source):Term: active customerProposed meaning: a customer with a transaction in the previous 90 daysBusiness context: transaction-based customer activity, separate from the current-contract definitionSource reference: customer activity guide, Section: “Active customer” Read status: read successfullyConflict status: noneReview the generated steps and run them over the selected documents. Compare a few extracted definitions with their sources, including a repeated term and one with different meanings in two business areas.If the flow merges those meanings, refine its consolidation instructions and rerun it. Any missing definition can be traced back through the extraction output and issue log.Read, extract, and log any exceptionsOnce the connectors are available and the workflow can access the document types in the chosen folders, the process can be tested. Include the actual mix you expect to use: PDFs, Google Docs and Slides, Excel workbooks or Word files.You may need to authorise a connector or a supported connector action before the workflow can read the relevant document type, but this is generally done only once. If you are sharing the workflow with your team, then they too will need access to the relevant documents.Ask the builder to log the source reference, read outcome, and any access, connection, or content problem for each selected source. Readable content proceeds to extraction; a failed read produces an issue for follow-up. Keep the issues in a separate log, not in the glossary file.Access problems and unsupported files will need to be addressed separately and then reprocessed if possible. For a file-format problem, a common, easy fix is just saving the unprocessed documents as a PDF.Otherwise, you can try to get your support team to enable a suitable connector or action in the Gemini Agent Flow Builder, if one is available, but maybe the PDF route could be a little quicker.Illustrative AI-generated flow. Successful reads continue through the generated steps; inaccessible sources and unresolved definitions enter the issue log.Consolidate meanings and remove duplicatesTake an illustrative “active customer” definition. One business unit may mean a customer with a transaction in the previous 90 days. Another may mean an account with a current contract.Describe that distinction in the consolidation prompt: compare the entity, business unit, time window, metric grain, and process, and retain the source reference. The builder creates the step that carries out that comparison. Repeated references to the same entity in the same context become one record, while the transaction-based and contract-based definitions remain separate.The extraction template supplies the details for that comparison. A second document repeating the 90-day transaction rule can contribute another source reference to the same definition. A document describing an active contract contributes to the contract-based definition.If two sources disagree about the time window for the same measure, the workflow records the conflict for the business owner to resolve.Bring an existing glossary or other business context into the same pass. Ask the builder to add steps that compare it with the new definitions, merge repeated meanings, check for inconsistencies, and deduplicate again. These steps run automatically over the extracted records and supplied glossary, retaining one entry per real entity and context.An old business case may use a definition that a current process document has replaced. Include an instruction to flag that discrepancy with both source references. The business owner can then confirm which definition applies, while the workflow has already consolidated the entries that agree.Share the workflow and bring the context togetherWorkflow sharing lets colleagues in other business areas run the workflow over folders they can access and contribute agreed definitions to the central glossary. Source permissions stay with the person running it. A user who hasn’t authorised a required connector can complete its OAuth authorisation when prompted during a run.The consolidated output can go straight to an agreed location or through a staging file, depending on your risk appetite and interpretation of YOLO. Make that destination an explicit choice. An extracted definition can still contain restricted information.Files uploaded to the agent’s Knowledge section are accessible to everyone who can use the shared agent, even if they can’t access the original files. Keep that distinction in mind when contributing documents or definitions to a shared location.Put the context to workThe consolidated definitions can feed data-model descriptions, measure definitions, and instructions for conversational systems. An analytics engineer can use proposed metadata alongside the model’s actual fields and calculations, with the source reference there when a definition needs checking.In Looker Conversational Analytics, LookML field names, labels, descriptions and types help map a business question to fields. The model supplies the joins, filters, aggregations, and access controls. A definition extracted from a process document can provide the business wording for a field description.For “active customer”, metadata attached to a transaction-based measure can carry the 90-day definition. A contract-based measure gets the current-contract meaning. A conversational system can then distinguish the measures, or ask for clarification, instead of mapping both to the same phrase.Run the extraction again as source records and business terms change. Supply the existing context for the merge, and use the issue log to follow up a newly readable file or a conflict that still needs an answer. Check changed meanings before replacing dependent metadata.This worked as a small-scale workflow. Across many business units, repositories, or a large data environment, source ownership, access, and conflicting definitions become a larger ingestion and maintenance job.If you are struggling to keep an updated business glossary for your data model and AI agents to use, or you just cannot get the business on board to spend the time to create one for you, maybe give this a try and let me know how you go.This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!Extracting Business Context the Easy Way was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.