Practical uses of Sherlock with My Portal on CloudFerro Cloud
Translating notices, reading long reports, and finding answers in technical documentation can take repeated manual work. Sherlock lets you try these tasks in a browser and use model requests in your own applications. You can begin with one document, then apply the same instructions to a collection of documents or several languages.
Sherlock is a managed generative AI service. You call a hosted model without deploying or operating that model yourself. My Portal also provides access to services for running your code, scheduling work, exchanging messages, and comparing results. The examples below show useful ways to combine those capabilities.
Translate into one or several languages
A language model can translate a notice, article, product description, or support reply. Give it the source text, the target language, and any terminology or formatting rules that matter. For technical documentation, ask it to preserve commands, filenames, links, and variable names while translating the surrounding explanation.
Programmatic translation uses the chat completion API. For example, your Python program can read an English service notice and request Polish, German, and French versions. One request per target language makes it easier to save, review, or retry each translation separately. This is a loop in your application using ordinary API requests.
Keep each result with its source document ID and target language. When the original changes, your application can submit the new version again. This is useful for recurring announcements and multilingual document collections. Test the selected model on the languages and specialist vocabulary you need, and review the translations before publishing them.
Summarize and adapt existing text
Sherlock can turn a long report into a short briefing, extract action items from meeting notes, or draft a reader-friendly explanation of technical material. Specify the audience and the desired result: for example, a five-point operational summary or a short update for users affected by maintenance.
For a data-processing workflow, supply the measurements and findings that your analysis has already produced and ask for a narrative summary. Keep those values available for comparison with the generated text. The benefit is a first draft based on your material, with less time spent repeatedly preparing the same kind of explanation.
Answer questions from your own documents
A team may need answers spread across manuals, procedures, and project reports. A Sherlock knowledge base holds the documents; a search agent retrieves relevant content; an assistant combines that content with a model and instructions to answer questions.
For example, a support assistant could help colleagues find a documented recovery procedure or explain a project convention. Start with a small, maintained collection using knowledge bases, then connect an agent and assistant through the linked guides. Check document-processing status before testing questions. This supplies document context for answers; it does not retrain the underlying model.
Turn scanned material into usable text
The OCR Playground extracts content from a PDF or supported image. Its output choices are Markdown, Text, and JSON. This is useful when a scanned page or screenshot contains text that you cannot copy directly.
After checking the extracted result, you can use the text as input for translation, summarization, or field extraction. For example, ask a language model to identify a report title, date, and project name using a defined set of fields. Validate the returned structure and values before storing them. The documented OCR example uses the web interface; the automated workflow later in this article starts with text that has already been extracted.
Interpret an image or explain code
When a suitable vision model is available to your project, you can ask a question about an image through the vision endpoint. Examples include describing a diagram or helping interpret a screenshot. OCR is useful for extracting written content, while a vision request lets you ask about the image in the context of a specific question. Check current model availability before choosing this approach.
For programming tasks, give a language model a short code fragment, an error message, and the intended behavior. It can propose an explanation, a change, or an example of how to process the data. Review and test generated code in your normal development environment before using it in a working service.
Connect Sherlock to other My Portal services
The My Portal product catalogue provides a common starting point for these services. The combinations below describe application designs you can build. Your code makes the API calls and moves the data between services; opening products in the same portal does not connect them automatically.
Service |
Useful role alongside Sherlock |
|---|---|
Try translation or classification in Python, inspect responses, and compare a small set of documents interactively before preparing a recurring job. |
|
Record model IDs, prompt versions, quality scores, and response times in MLflow. The Sherlock comparison guide supplies a working example of direct API calls and experiment logging. |
|
Schedule a Python workflow to read new text, request translations or summaries, check the results, and save them. Put the Sherlock calls and handling of failures in your workflow tasks. |
|
Implement a short request handler that receives text, calls Sherlock, and returns a result. Check execution time limits; use background workers for longer jobs. |
|
Queue translation or classification jobs until workers can process them. A worker consumes a job, calls Sherlock, and stores the result; the broker transports messages. |
|
Run an application backend or persistent workers in containers. These programs call Sherlock for model inference while your deployment handles requests, storage, and job state. |
Each service has its own access requirements and connection settings. A Sherlock project API key authorizes Sherlock calls; AI Experiments, a broker, or storage requires its own credentials and permissions. Keep the Sherlock key in the backend’s environment or secret storage, outside source code and browser-delivered JavaScript. Confirm the selected organization, applicable regions, and network access when preparing a combination.
Example: a recurring multilingual bulletin
Suppose your team prepares an English operational bulletin and needs reviewed versions in Polish, German, and French. A useful application design is:
Read the source. A scheduled Workflows task loads the latest approved text and its document ID.
Request translations. A Python task sends one Sherlock request for each target language, using the same terminology and formatting instructions.
Check and save results. The application checks that an answer was returned, checks required identifiers and links, and stores each draft with its language, source version, model ID, and prompt version. A reviewer checks meaning and terminology before publication.
Track the outcome. The application records response times and review scores in AI Experiments so later model or prompt changes can be compared against the same sample texts.
For a first trial, a notebook and Sherlock are enough. Add Workflows when the job needs a schedule. If submissions arrive in bursts, add Message Broker and workers to hold and process pending jobs. These are optional additions to meet the application’s workload.
Start with a task you can evaluate
Choose a small set of representative inputs and decide what a useful answer must contain. Translation needs review of meaning and terminology; classification can be checked against expected labels; extraction can be checked for missing or incorrect fields. Compare models and prompts on the same inputs before expanding the workload.
Use Sherlock’s consumption and event views to inspect project activity and input/output token counts. Record quality separately. More target languages or repeated requests consume more model usage, and the notebook, workflow, or worker hosting your code has its own resource usage.
What to do next
Prepare account and project access through My Portal.
Try a representative prompt and compare answers in the Chat Playground.
Follow Get started with the Sherlock API when moving the task into Python or another application.
Choose a worked example from Examples and integrations to compare models, translate queued text, or run a translation function.