This is a step-by-step guide to create a Corpus and upload resources to it.
First, log in to Glob.AI OS. In the Navigation header, select the Organization and Project you want to work with. Then, from the Hub selector, select Agentic Hub.
Next, on the left side of the screen, locate the Side Menu. In this menu, select Build > Corpus.
The Corpus Dashboard opens in the central working area.
On the Corpus Dashboard, select + New.
In the Create New Corpus dialog, complete the following fields:
- Type: Select the type of corpus, such as Vector Store.
- Name: Enter a descriptive name for the corpus, such as Documentation Corpus.
- Description: Enter a description of what the corpus is for, such as Indexed internal documentation used to answer users' research requests. Keep the description free of line breaks and special characters. Otherwise, saving may fail.
Select Save. The new corpus appears on the Corpus Dashboard, along with its file count, status, and total storage used.
On the Corpus Dashboard, locate the corpus card and select Edit.
The corpus detail view opens, showing the following tabs: Configuration, Embeddings, Chunking, and Ingestion. On the Configuration tab, you can update the Corpus Name, Description, Type, and Active status, and this is also where you find the Resources section.
In the Resources section, select + New Resource.
You can upload a resource using one of the following two methods:
- Default Upload: A quick upload using the corpus's default settings for provider, embeddings, and chunking.
- Advanced Upload: A guided, multi-step upload that lets you configure the provider, metadata, embeddings, chunking, and ingestion settings for the resource.
To begin, select the corresponding option:
- Leave the Advanced Settings toggle off to use the default upload.
- Turn on the Advanced Settings toggle to use the advanced upload.
With Advanced Settings off, the dialog shows only two steps: Select and Finish.
- On the Select step, select Add File, or drag and drop a file into the upload area.
- The maximum file size is 100 MB.
- Supported file types include txt, pdf, doc, docx, ppt, pptx, and xls, among others.
- Select Upload.
- Wait for each file to show a success status.
- Select Close, or select Upload more to add additional resources.
With Advanced Settings on, the dialog expands into a six-step:
- Upload
- Metadata
- Embeddings
- Chunking
- Ingestion
- Finish
Use Advanced Settings when you need to customize how a resource is processed. If you don't enable it, the corpus uses the default provider, embeddings, and chunking settings.
A step progress indicator at the top of the dialog reflects these six steps and marks each completed step as done as you move forward.
Select the Provider, such as Globant Enterprise AI, and add the file, or drag and drop it into the upload area.
The Continue button is enabled only after you upload at least one file. Once you add a file, the Cancel button also becomes visible and active.
Select Continue.
Add custom key-value pairs to associate with the resource. For example, when uploading a text-based document with tables, such as an HTML page listing supported file formats, you could use the following values:
- Select + Add line.
- Enter a Key, such as source, and its corresponding Value, such as wiki.genexus.com.
- Repeat for each additional metadata pair you need, such as category / RAG File Formats.
- Select Continue.
From this step through Ingestion (steps 2 to 5), you can select Continue, Cancel, or Back to navigate. Selecting Continue on the Ingestion step (step 5) uploads the files completely and takes you to the Finish step (step 6).
Configure how the resource is embedded for semantic search. For example, you could use the following values:
- Provider: The embedding provider, such as openai.
- Model Name: The embedding model, such as text-embedding-3-small.
- Dimensions: The vector size used by the model, such as 1536.
- Type: An optional embedding type. Leave it empty.
- Use Proxy: Route embedding requests through a proxy. Enable it.
- Cache: Cache embedding results to avoid reprocessing. Enable it.
- Batch Size: The number of items processed per embedding request, such as 100.
- Mode: An optional embedding mode. Leave the default.
Select Continue.
Configure how the resource is split into chunks before indexing. For example, you could use the following values:
- Chunk Size: The maximum size of each chunk, such as 1000.
- Chunk Overlap: The number of overlapping units between consecutive chunks, such as 100.
- Parent Document: Enable this option to preserve the parent document structure. Leave it disabled.
- Chunk Strategy: The strategy used to split the content, such as byLayoutType, since the document contains structured tables.
Select Continue.
Configure how the resource content is extracted and processed. For example, you could use the following values:
- LLM Provider: The provider used for ingestion, such as openai.
- Model Name: The model used for ingestion, such as gpt-4.1-mini.
- Strategy: The ingestion strategy, such as auto.
- Structure: An optional document structure setting. Leave it empty.
- Start Page / End Page: The page range to ingest. Leave both empty to ingest the full document.
- Password: The password for protected files, if applicable. Leave it empty, since the document isn't protected.
- Logo Process: Enable this option to process logos found in the resource. Leave it disabled, since the document has no logos.
- Dialogue: Enable this option to process dialogue content. Leave it disabled, since the document has no dialogue.
- Image Prompt: The prompt used to describe images found in the resource.
- Scanned Prompt: The prompt used to extract text from scanned pages.
- Table Prompt: The prompt used to extract and structure tabular data.
- Media Prompt: The prompt used to generate captions for video frames.
- Whisper Model: The model used to transcribe audio, such as tiny. Leave the default, since the document has no audio.
- Merge: An optional setting to merge extracted content. Leave it empty.
- Frame Sampling Rate: The rate at which video frames are sampled. Leave it empty, since the document has no video.
Select Continue.
The Close button appears only on this step. It remains inactive until all uploaded files finish processing, at which point it becomes active.
A progress bar shows the upload progress for each file individually.
Review the confirmation message, and wait for each file to show a success status. Select Close, or select Upload more to add additional resources.
Once you have created a corpus, you can use it in an Agent. Go to Build > Agents, and create a new Agent following the steps in How to create an Agent.
In the Configuration Tab of an Agent, add Background and Behavior guidance to the Agent's prompt. For example, ask the Agent to:
- Be concrete.
- Stay to the point.
- When grounding links are provided, include the associated reference as part of the response, so the Agent cites the source directly.
In the AI & Tools Tab of an Agent, in the Tools and Agents section, select + Add to open the Select Tools and Agents dialog. This step is essential, since without a correctly configured com.globant.geai.semanticsimilarity.retrieve Tool, the Agent can't reach the corpus and answers using its base model only.
In the Select Tools and Agents dialog, expand the Semantic Similarity tools, and select the checkbox next to com.globant.geai.semanticsimilarity.retrieve. Then, select the Tool in the list to view and configure it in the right pane.
The Tool configuration is organized into three tabs:
- Corpus selection: Use the Select Corpus field to choose the corpus the Agent should query, such as Documentation Corpus. If the Tool doesn't have a corpus selected, a warning indicates that the Corpus is not configured.
- Retrieval Settings: This tab is divided into two parts:
- The top section sets the Score threshold (a float value between -1 and 1; an error message appears if you enter a value outside this range) and the Chunk count.
- The Rerank section lets you set a Provider (OpenAI, Cohere, or AWS Bedrock), a Model (populated based on the selected Provider), a Top-K (a positive integer), and a Relevance score (a value between 0 and 1, with support for multiple decimal places).
- Filters: Select + Add line to add a filter row, then complete its Key, Operator (=, !=, >, >=, <, <=), and Value. Use the trash icon to remove a row.
Select Save to add the Tool to the Agent.
These settings are saved as part of the Agent's configuration when you save the Agent. The Tool itself does not store the retrieval settings.
Save the Agent, then test it by asking a question that the uploaded resources can answer, such as a question about the content of a specific file. A correctly configured Agent calls the corpus Tool, retrieves the relevant chunks, and answers the question, including source references if you configured it to do so. Once you've tested the Agent, you can Publishing an AI Solution.
Since version 2026-06.