Information Condensation Techniques for Prompts

An agent's short-term memory, primarily managed within its context window, is a finite resource. When dealing with extensive interactions, lengthy documents, or detailed tool outputs, the raw information can quickly overwhelm this limited space. Information condensation techniques are therefore indispensable. By intelligently reducing the volume of information while preserving its essential meaning, agents can handle more complex scenarios, maintain longer dialogues, and operate more efficiently. Effective strategies for using prompts to guide an LLM in condensing information for an agent's memory are presented.

The core idea is to transform verbose information into a more compact form that the agent can readily use or store. This isn't just about making things shorter; it's about retaining the most relevant details for the agent's current objectives and future actions.

Prompting for Summarization

One of the most direct ways to condense information is through summarization. You can instruct the LLM to create a concise version of a given text. Your prompts can guide the style, length, and focus of these summaries.

Abstractive vs. Extractive Summarization Cues

While you don't typically specify "abstractive" or "extractive" in a prompt, your wording can lead the LLM towards one or the other:

• To encourage abstractive summarization (where the LLM generates new phrasing to capture the essence), use prompts like:

  • "Summarize the following customer feedback in one paragraph, highlighting their main sentiment and suggestions."
  • "Provide a brief overview of the attached article, focusing on its key findings."
  • "Condense the last three turns of our conversation into a single sentence that captures the current state of the request."

• To encourage extractive summarization (where the LLM pulls key sentences or phrases directly from the text), you might ask:

  • "Extract the most critical sentences from this report that outline the project risks."
  • "Identify the main points from the user's query." (This often results in extracted phrases).

In practice, LLMs often produce a hybrid. The important part is that your prompt clearly defines the desired outcome and the information to prioritize.

Use Cases for Summarization in Agents:

• Conversation History: Condensing prior turns of a conversation to keep the agent aware of context without exceeding token limits. For example, after every few turns: "Summarize our conversation so far, focusing on decisions made and open questions."
• Tool Outputs: Many tools return verbose JSON or lengthy text. An agent can be prompted to summarize this output: "The search tool returned the following: [long text]. Summarize the relevant findings for answering the user's question about X."
• Document Processing: When an agent needs to "read" a long document, it can process it in chunks, summarizing each chunk and then summarizing the summaries to build a comprehensive understanding.

Keyword and Keyphrase Extraction

Sometimes, a full summary isn't needed, but rather a set of significant terms that represent the core topics of a piece of text. Prompting for keywords or keyphrases can be a very efficient condensation method.

• Example Prompts:
  • "Extract the 5 most important keywords from the following text: [text]."
  • "List the key technical terms mentioned in this document."
  • "What are the main topics discussed in this user request?"

Keywords can be used as tags, for quick relevance checks, or as input for further information retrieval from a knowledge base.

Structured Data Conversion

Often, information spread across natural language can be more effectively stored and processed if converted into a structured format like JSON. This is a powerful condensation technique because it not only reduces verbosity but also organizes the information for easier programmatic access by the agent or other systems.

• Example Prompts:
  • "Convert the following user request into a JSON object. The request is: 'I want to order 3 large pepperoni pizzas and a diet coke for delivery to 123 Main St.' The JSON should have keys 'items' (an array of objects with 'name' and 'quantity'), 'drink', and 'address'."
  • "Analyze this product review and extract the product name, the rating (1-5 stars), and a list of pros and cons mentioned. Present this as a JSON object."
  • "From the following meeting transcript snippet, identify action items. For each action item, provide the item itself and the person assigned, formatted as a list of JSON objects: [transcript snippet]."

This method is particularly useful for extracting specific pieces of information from unstructured text and making them directly usable by the agent's logic or tools.

Guiding the Condensation Process

Effective condensation relies on clear instructions within your prompts.

• Length and Detail: Be explicit about the desired conciseness.

  • "Summarize in a single sentence."
  • "Provide a bullet-point list of no more than 3 items."
  • "Condense this to under 100 words."

• Focus and Relevance: Direct the LLM on what information is important to retain, especially in relation to the agent's current task or overall goal.

  • "Summarize the user's problem description, focusing on the symptoms they are experiencing, and ignore any pleasantries."
  • "From this API response, extract only the 'id' and 'status' fields."

• Maintaining Integrity: While brevity is good, ensure critical information isn't lost.

  • "Condense the previous steps, but ensure all constraints and user preferences mentioned are retained."

Progressive Condensation

For ongoing interactions or tasks that evolve over time, progressive condensation is a valuable strategy. Instead of re-processing all historical data, the agent maintains a running summary or a condensed state that is updated with new information.

• Prompting Strategy:

  1. Provide the current condensed summary/state.
  2. Provide the new piece of information (e.g., latest user message, new tool output).
  3. Instruct the LLM to integrate the new information into the existing summary, creating an updated, still condensed, version.
  • Example: "Here's a summary of our progress: [previous summary]. The user just added: [new user message]. Update the summary to reflect this new information, keeping it concise."

This approach helps manage context effectively in long-running agentic workflows.

The following diagram illustrates the general flow of information condensation guided by prompts:

An agent uses a specific prompt to instruct the LLM to process raw information and produce a condensed version, which is then used to update the agent's working memory or inform subsequent prompts.

Balancing Brevity with Detail

The primary trade-off in information condensation is between conciseness and information fidelity. Overly aggressive condensation can lead to the loss of important details, hindering the agent's ability to perform its tasks accurately. Conversely, insufficient condensation fails to alleviate context window pressure.

• Iterative Refinement: Finding the right balance often requires experimentation. Test different condensation prompts and observe their impact on the agent's performance and the quality of its memory.
• Task-Dependent Condensation: The optimal level of detail in a condensed summary can vary depending on the agent's current task. For instance, a summary for a quick status update might be shorter than one intended to capture all details for a complex decision-making process.

By mastering these prompt-driven condensation techniques, you equip your AI agents to handle larger volumes of information more effectively, leading to improved coherence, better decision-making over extended interactions, and more efficient use of the LLM's capabilities. This condensed information is not only useful for short-term recall but also serves as an excellent candidate for storage in long-term memory systems, which we will explore further.