When interacting with Lampi AI in a chat, an AI assistant is always selected by default. However, you have the flexibility to manage and choose an AI assistant that is best suited for your tasks. This capability allows you to optimize your interactions and results based on the unique requirements of your projects.
Having specialized AI assistants for different tasks can greatly enhance your workflow. Each assistant can be designed with specific roles and responsibilities, making it easier to access the right expertise when needed. This tailored approach not only improves efficiency but also ensures that the responses you receive are more relevant and accurate to your specific needs.
How to create your custom AI assistant?
To create your own custom assistant, open the page "Settings" and follow these steps:
1. Define the name of your assistant:
Choose a name for your assistant, such as “Financial Analyst” or “Customer Support Specialist.”
2. Enter a System Prompt and instructions
The system prompt provides general instructions that guide the model's behavior. This is where you will set the foundational instructions for your assistant's behavior.
System Prompt and Prompt: When you interact with Lampi AI, there is always two inputs, a "system" prompt that contains instruction for the model to follow and a "prompt" input with various data sources (e.g., the user message, context, data sources, etc.) that you enter in a chat.
- Give your assistant a role: You can dramatically improve your assistant performance by giving it a role. The right role can turn your assistant from a general assistant into a domain expert. In complex scenarios like legal analysis or financial modeling, role prompting can significantly boost your assistant performance and adjust communication and written style.
-
Provide detailed instructions, including:
- Tone of communication: Specify how the assistant should express itself (e.g., "Respond in a professional manner").
- Data source references: Indicate how the assistant should utilize different data sources depending on the inquiry. To make your assistant focus on only the context retrieved from different tools, mention that the assistant should use only the context it has access to and if it doesn't know the answer to reply "I'm sorry, but I don't have enough information to answer."
- Company-specific Information: Provide context or reference material relevant to your situation (e.g., “Use the following mail link”).
- Response Format: If you need a specific output structure, clarify this in your instructions.
- Specific text that you want the model to output in certain situations.
- Reasoning approach: Think through the steps required for the task and encode this into the system prompt.
Note that you can add some instructions in the section "Extended User Prompt".
Content Writer Assistant:
You are an expert content writer assistant, skilled in supporting writers with creating engaging, original, and impactful content.
You can assist in brainstorming ideas, developing outlines, crafting compelling introductions and conclusions, and refining content for clarity and coherence. You are knowledgeable about SEO, audience engagement strategies, and different writing styles to match the specific needs of various types of content. You have extensive experience in writing SEO-optimized blog posts.
Use an engaging and informative tone of voice. Format your response using markdown and structure your responses with clear tips and suggestions to make them easy to implement.
If you are uncertain about your answer, clearly state that the information is not available and ask the user to clarify or provide additional details.
Financial Analyst:
You are an experienced financial analyst, specializing in providing detailed, insightful, and data-driven information.
You can assist with financial analysis, investment evaluation, and risk management. Your expertise extends to understanding economic trends, financial ratios, and data interpretation to provide comprehensive insights and strategies.
Your responses should be accurate, data-driven, and align with best financial practices. Every answer should be considered a reliable source for informed decision-making.
Adopt a formal and analytical tone. Format your response using markdown and organize your responses with structured paragraphs to ensure clarity.
If you are uncertain about your answer, clearly state that the information is not available and ask the user to clarify or provide additional details.
Grammar Assistant:
You are an expert grammar assistant dedicated to enhancing the quality and precision of written communication.
Your task is to take the text provided and rewrite it into a clear, grammatically correct version while preserving the original meaning as closely as possible. Correct grammatical and punctuation errors, spelling mistakes, verb tense issues, suggesting stylistic improvements, and ensuring clarity and conciseness. You are well-versed in grammar rules, punctuation, and syntax.
Each answer should contribute to producing polished, error-free text.
Always start with the corrected sentence and continue with a very concise note on the changes made and why.
Translation Assistant:
You are an expert translation assistant focused on providing accurate, context-aware translations between languages.
You can support with translating documents, maintaining cultural nuances, and suggesting language adjustments for clarity and tone. You have a deep understanding of idiomatic expressions, formal and informal registers, and technical terminology across multiple languages.
Your answers should be accurate and maintain the intended meaning of the original text. Each response should be a trusted source for high-quality translation support.
3. Add capabilities
To improve the capabilities of your AI agent, you should consider two possibilities:
Reranker
Reranker might improve the results of the Search tool (i.e., when you select documents to interact with).
Reranking in Retrieval-Augmented Generation (RAG) refers to the process of reordering or refining a set of initially retrieved information or insights (i.e., chunks) based on their relevance to a query.
When you interact with files, a retrieval system will extract all the information that appear to be relevant for the query. However, the first batch of documents or insights retrieved can be broad, including both highly relevant and somewhat tangential information. Reranking addresses this by filtering and reordering these documents according to refined relevance scores, reducing noise and irrelevant information.
Mode selector
The Mode Selector is another way to improve the quality of the Search tool.
When you have selected the Mode Selector, each time you enter a prompt, the system will analyze it to determine it requires advanced agentic capabiilities, such as iteration, or not.
- If the system considers your question is "simple" and does not require extended capabilities (e.g., "Tell me the name of the founders of the company?"), the system will execute the Search function.
- If the system considers your question is more complex, a more advanced system will be activated to answer the query: the AI assistant will analyze documents page by page and document by document with iterative behavior. Each relevant block of context is saved and enriches the overall context of the query. This system is particularly well suited to long, detailed queries requiring cross-analysis or complex synthesis. For example, it can handle queries involving several levels of comparison or advanced searches.
Warning: Depending on the size of the files you selected, the duration of the task can be particularly long. It will be treated as a task and you will be able to track its status in the Tasks page. Activate the Mode Selector only if you understand perfectly how to use it.
Keyword
The mode Keyword applies only with the Search tool and involves agentic process where each of your query is processed and rewritten to maximize the AI outputs.
The system rewrites the query in multiple similar ways and perform a different search for all re-written queries. This step helps to maximize the search engine.
Warning: When selected, this option will apply to all your queries with the Search tool. As such, each of your query will take more time to process. Only use it for specific use cases.
3. Select the model and its settings
Model (LLM)
You can choose the model that will power your assistant and define its settings, including:
-
LampiLLM r1
-
LampiLLM ZB
-
Mistral Instruct v0.2
-
Azure GPT4-o-mini 128k
-
Phi3Vision Instruct 128k
-
Nemotron 70b 128k
-
Nvidia 405b 128k
-
Azure GPT4-o 128k
-
Cohere
-
Claude 3
-
DeepSeek V3 0324
-
Llama 4
-
Qwen2.5 32B VL (SW)
-
Gemini Flash 2
-
Gemini Flash 2.5
Note: Lampi AI is designed to provide a secure and confidential way to interact with AI and your knowledge. When you select a model hosted by your company or Lampi, the data you are providing to the model stays secure and under control. However, if you select a proprietary model, such as GPT-4, some data will flow to the model provider to process your request. As a result, we may not have full visibility or control over how the provider handles the data you chose to process with them.
The best model depends on your use case. Make sure to evaluate the performance of the model you select on your task.
Once you have selected the model, here are some parameters to consider:
Temperature
Temperature controls the diversity of LLM generation. You can adjust the temperature settings for your models (for chat interaction, summary, and AI agents).
Temperature in LLMs is the trade-off between reasoning and creativity.
- Low temperature: high reasoning & low creativity
- High temperature: low reasoning & high creativity
The optimal temperature doesn't exist. It depends on the tasks and use cases.
Low Temperature
Close to 0
- High reasoning
- Low creativity
- Predictable output
- Conservative responses
High Temperature
Close to 1
- Low reasoning
- High creativity
- Diverse output
- Unexpected responses
Use Cases
Low Temperature
- Generating factual content
- Answering questions
- Translations
High Temperature
- Brainstorming
- Creative writing
- Generating diverse responses
Tips: If you intend to use the Search tool with your AI assistant, select a temperature close to 0.
Max generation token
Each model has a limit max number of input and output tokens. To adjust the limit for a particular model you can modify the max tokens parameter. This parameter is an easy way to control the response lenght. However, max tokens actually cuts off the response when it reaches the limit. So using max tokens is very practical and recommended for high-usage applications.
Note: You cannot increase the max tokens beyond the maximum supported for a particular model. To generate longer answers such as reports, you can use AI agents.