Volatile Knowledge

Introduction

The Volatile Knowledge module allows agents to use files attached by final users on an instance. Serenity* Star automatically analyzes and segments these documents, creates and stores embeddings, and will then use the information obtained throughout the execution.

How to Attach a file

The method for attaching a file to your execution will depend on the channel you are using.

API

To attach a file when executing an agent through the API, you first need to upload the file using the Volatile Knowledge endpoint. Once the file was processed, it can be attached in the Execution endpoint.

Preview

Use the attach file button to attach a file to a message or instance.

WhatsApp

If you are using our WhatsApp integration, you can attach files using the WhatsApp interface and they will be directly processed.

Volatile Knowledge Strategy

The Volatile Knowledge Strategy controls how the agent processes and incorporates files uploaded by users during a conversation. Unlike knowledge base files — which are pre-configured per agent — volatile files are transient and session-specific.

To configure the strategy:

1. Open the agent in the Agent Designer and click the Knowledge card.

   

2. Under Volatile Knowledge, select the strategy that best fits your use case.

   

Attach Full File

The agent includes the complete content of each uploaded file in the conversation context and sends it to the model on every message. If the total content of all attached files exceeds the available context limit, a validation error is returned and the message is not sent.

How it works:

1. The content of all uploaded files is extracted and combined.
2. The total size is checked against the available context limit.
3. If it fits, the full content is included in the context sent to the model.
4. If it exceeds the limit, an error is returned and the message is not sent.

When to use it:

• Files are short and the model needs full visibility of their content to respond accurately.
• The task requires precise, verbatim information — partial retrieval is not acceptable.
• You want simple, deterministic behavior with no retrieval logic involved.

Limitations

Not suitable for large files or conversations with many attachments. If the combined file content exceeds the context limit, an error is returned and the user must reduce the amount of attached content before retrying.

Similarity Search

Instead of including full file content, the agent performs a semantic search on every message to retrieve only the sections most relevant to the user's query.

How it works:

1. Each uploaded file is indexed for semantic search.
2. The user's message is used as the search query.
3. The most relevant sections across all files are retrieved and included in the context.

When to use it:

• Files are large and including their full content would exceed the context limit.
• User questions are specific and only a portion of each file is relevant at a time.
• You want predictable context usage regardless of file size.

Limitations

A text query from the user is required to perform the semantic search. If the message contains no text (e.g., a voice message without transcription), the files are skipped entirely.

Hybrid

The agent first evaluates whether the full content of all uploaded files fits within the context limit. If it does, the complete content is used. If not, it automatically falls back to semantic search and retrieves only the most relevant sections.

How it works:

1. The total size of all uploaded files is evaluated against the context limit.
2. If the content fits, the full file content is included in the context.
3. If it exceeds the limit, a semantic search is performed and only the most relevant sections are included.

When to use it:

• File sizes are unpredictable and you want the best outcome in both scenarios: full content when it fits, targeted retrieval when it does not.
• You prefer an automatic fallback over returning an error to the user.

Limitations

If the content exceeds the context limit and the user message contains no text, files are not included in the response at all.

Agentic File Processing

The agent decides autonomously whether to read uploaded files and which ones to access, based on the conversation context. Rather than including file content on every message, the agent evaluates each turn and retrieves file content only when it determines it is necessary to answer accurately.

How it works:

1. The agent receives the list of available uploaded files along with the conversation.
2. It evaluates whether any file content is needed to answer the current message.
3. If so, it selects the relevant files and retrieves their content.
4. The retrieved content is used to compose the response.

When to use it:

• Conversations involve multiple uploaded files but only a subset is relevant to each question.
• You want to avoid unnecessary file processing on messages that do not require it.
• Users reference different files across turns and selective retrieval improves response quality.

Considerations

The agent may occasionally decide not to read a file even when doing so would improve the response. Retrieval accuracy depends on the reasoning capabilities of the configured AI model.

Supported File Formats

The following file formats are supported for upload and processing in the Volatile Knowledge module:

FILE FORMAT	MIME TYPE
.txt	text/plain
.pdf	application/pdf
.doc	application/msword
.docx	application/vnd.openxmlformats-officedocument.wordprocessingml.document
.csv	text/csv
.md	text/markdown
.jpg, .jpeg	image/jpeg
.png	image/png