Skip to main content

Chat Best Practices

This comprehensive guide provides practical tips and strategies for getting the most out of your interactions with Galene.AI, whether you're using the general Galene.AI Chat or working with specialized agents.

Understanding AI Model Knowledge​

The AI models in Galene.AI have broad, generalist knowledge spanning many topics, but this knowledge is frozen at the time of training and may not include the most current information or specialized domain knowledge.

What Models Know Best:

  • General concepts, definitions, and common knowledge
  • Widely documented topics and established principles
  • Standard processes and best practices across industries
  • Technical fundamentals and programming concepts

What Models Don't Know:

  • Your organization's specific data, policies, or procedures
  • Niche or highly specialized domain knowledge
  • Current events and real-time information (unless web search is enabled)
  • Recent developments after their training date
  • Your personal context, role, or background
Knowledge Augmentation Approach

Galene.AI uses knowledge augmentation rather than relying solely on pretrained model knowledge. This means:

  • File uploads, agent configurations, personalizations and knowledge bases provide the AI with your specific information
  • Web search enables access to current, real-time data
  • Database connectors give agents access to organizational data (see Connecting to Enterprise Data)

Why this matters: Augmenting knowledge is more reliable, verifiable, and cost-effective than retraining models. It personalizes the AI to your needs without the risks and expenses of model training.

Best Practice: For specialized questions and verifiable answers, always provide relevant documents, enable web search for current information, or use agents with configured knowledge bases rather than relying on the model's pretrained knowledge alone.

Model Selection Strategy​

Reasoning vs Fast Model: Making the Right Choice​

Choosing the appropriate model significantly impacts the quality and reliability of your results.

Use the Reasoning Model When:

  • Working with complex queries that require deep analysis
  • Using tools like web search, file analysis, or database queries
  • Accuracy and thoroughness are more important than speed
  • Dealing with high-value information that needs verification
  • Performing multi-step problem solving
  • Working on tasks where errors would be costly

Use the Fast Model When:

  • Asking simple, straightforward questions
  • Speed is essential and errors are easily caught
  • Performing quick iterations on creative content
  • Generating drafts that will be heavily edited
  • Working with easily verifiable information
Fast Model Limitations

The Fast Model prioritizes speed over accuracy. It is:

  • More likely to make small mistakes in complex reasoning
  • More prone to confusion with multi-layered questions
  • Higher risk of hallucination when dealing with facts or data
  • Less reliable for information that's difficult to verify immediately

Recommendation: Use the Reasoning Model as your default, especially for important tasks. Reserve the Fast Model for simple queries where you can quickly verify the output.

Switching Models Mid-Conversation​

Recommendation: Avoid switching models during a conversation.

Why:

  • Different models may interpret the conversation context differently
  • Response style and depth may become inconsistent
  • The conversation flow may feel disjointed
  • Results may be less predictable

Better Approach: If you need to switch models, consider starting a new conversation with the model better suited to your current needs. You can reference key points from the previous conversation if needed.

Crafting Effective Messages​

Message Structure and Clarity​

Clear, well-structured messages lead to better AI responses. The AI doesn't have access to unspoken context or assumptions - everything must be explicit.

Essential Elements:

  • Be specific about what you want
  • State your context and role
  • Specify your desired outcome
  • Define format and structure
Examples of Effective vs. Ineffective Messages

Be Specific About What You Want

  • ❌ "Tell me about this document"
  • ✅ "Summarize the key findings from this document in 5 bullet points"

State Your Context and Role

  • ❌ "How should we approach this project?"
  • ✅ "I'm a project manager at a B2B SaaS company. How should we approach launching a new enterprise feature?"

Specify Your Desired Outcome

  • ❌ "Help me with marketing"
  • ✅ "Create a 3-month content calendar for LinkedIn posts targeting CTO-level decision makers in fintech"

Define Format and Structure

  • ❌ "Give me information about X"
  • ✅ "Provide a comparison table showing the pros and cons of approaches A, B, and C"
What the AI Doesn't Know (Unless You Tell It)

Unless instructed via personalization, knowledge base, or agent configuration, the AI doesn't know:

  • Your job title, department, or industry expertise level
  • Your company's products, services, or market position
  • Your personal preferences or communication style
  • The context behind your question or project

Examples of providing context:

  • "Provide a highly technical explanation suitable for engineers"
  • "Assume I'm familiar with basic Python but new to machine learning"
  • "In the context of our company operating in the healthcare sector..."
  • "I'm preparing this for executive leadership, so focus on business impact"
Ensuring the Agent Has What It Needs

Before sending your message, verify the agent has access to:

  • What you want accomplished
  • Why and the surrounding context
  • Who this is for and the appropriate expertise level
  • The format you need (list, table, paragraph, etc.)
  • Enough specificity to provide a targeted response

Remember: this information may already be available through your agent's configuration, uploaded files, or other sources—but if not, include it in your message.

RAG and Document Analysis​

When working with uploaded files or agent knowledge bases, understanding how Retrieval Augmented Generation (RAG) works helps you write better questions.

Quick Overview: RAG uses semantic search to find relevant content based on meaning, not just keywords.

How RAG Works - Technical Details

Galene.AI uses semantic search to find relevant content in your documents based on the meaning of your question, not just keyword matching. The system:

  1. Chunks documents into meaningful segments
  2. Creates summaries to provide context
  3. Uses both text and visual analysis for comprehensive understanding
  4. Retrieves the most relevant sections based on your query
  5. Provides those sections to the AI for answering

For a complete technical explanation, see Adding Knowledge Sources.

What RAG Does Well vs. What It Struggles With

What RAG Excels At:

✅ Vertical, in-depth analysis - Explaining complex concepts within documents

✅ Conceptual understanding - Finding information based on meaning

✅ Context-rich answers - Understanding relationships between ideas

✅ Handling long documents - Processing documents of any length

✅ Multi-format comprehension - Analyzing text, tables, and visual layouts together

What RAG Struggles With:

❌ Programmatic requests - "How many times is the word 'revenue' mentioned?"

❌ Complete enumeration - "List every reference to this concept" (context capacity limits may retrieve only a subset)

❌ Cross-document comparisons - "Compare all sections across 10 documents" (ask for specific sections instead)

❌ Transversal statistical analysis - "Calculate averages across all files"

Troubleshooting: When AI Can't Find Information

If the AI says it cannot find information in your documents:

  1. Verify the file is uploaded and processed

    • Check the file management sidebar or Agent's configuration
    • Ensure status shows "Processed"

  2. Make your query more specific

    • Include relevant keywords and context
    • Use explicit terminology from the document

  3. Rephrase with explicit details

    • Instead of "the topic mentioned earlier", state the topic explicitly

  4. Break down complex questions

    • Ask about one specific aspect at a time

Example of Improving Search:

❌ "What does it say about the timeline?"

✅ "What is the project timeline for Phase 2 of the database migration mentioned in the technical specifications?"

Effective Document Questions

General → Specific: Move from broad to focused questions

  • First: "What are the main topics covered in this document?"
  • Then: "Explain the security protocols described in section 3"
  • Finally: "What specific authentication methods are recommended?"

Use Explicit References: Help the semantic search find the right context

  • Instead of: "Tell me more about what you mentioned"
  • Write: "Provide more details about the API rate limits mentioned in the integration guide"

Providing Context​

Key Point: The AI only knows what you explicitly tell it. While it may adapt to your writing style, it cannot infer your background, role, or organizational context.

Best Practice: Front-load context in your first message - the AI retains it throughout the conversation.

Context Examples for Different Scenarios

Professional Context:

"I'm a senior data analyst at a retail company analyzing customer churn.
I need to present findings to non-technical stakeholders..."

Technical Level:

"I'm familiar with basic SQL but haven't worked with window functions.
Can you explain this query with examples?"

Organizational Context:

"In our company, 'Q3' refers to October-December (fiscal year starts in April).
When analyzing this financial data..."

Purpose and Audience:

"This documentation will be used by external developers integrating with our API.
They're experienced programmers but new to our platform..."

Document Search Mode​

The Document Search mode selector lets you control how aggressively the assistant searches your documents before answering. Choosing the right mode for your task improves both answer quality and response speed.

ModeBehaviorBest For
AutomaticThe system decides when and how to searchGeneral use — suitable for most interactions
OffSkips document search entirelyOperational tasks (for example tool usage) where document context is irrelevant
StrictSearches with strict criteria; does not broaden the search if no reliable results are foundHigh-stakes questions where a "no result" is preferable to an uncertain answer
RelaxedSearches broadly and retrieves more content, accepting lower precisionExploratory questions where coverage matters more than exactness

Practical guidance:

  • Automatic works well for the vast majority of conversations — leave it as-is unless you have a specific reason to change it.
  • Switch to Off when asking about platform features or any request where searching your documents would not help.
  • Use Strict when working with any material where an incorrect retrieval would be worse than no retrieval.
  • Use Relaxed when brainstorming or exploring a topic and you want the AI to surface loosely related content that might spark ideas.
Combining Modes with Good Prompts

Even in Relaxed mode, a specific, well-formed question produces better search results than a vague one. Mode selection and prompt quality are complementary — the mode sets the retrieval threshold, but the query itself guides what gets retrieved.

Follow-up Questions and RAG​

When asking follow-up questions about documents, remember that each query triggers a new semantic search in your knowledge base.

The Challenge:

When you write: "As you mentioned before in point 2..."

  • The AI understands what you're referring to from the conversation history
  • But the semantic search for documents may not retrieve the same context based on this vague reference
  • This can result in less relevant document sections being retrieved

The Solution:

Make follow-up questions self-contained with explicit context:

❌ Vague Reference:

You: What are the main security features?
AI: [Lists 5 security features]
You: Can you elaborate on point 2?

✅ Explicit Reference:

You: What are the main security features?
AI: [Lists 5 security features including "End-to-end encryption"]
You: Can you provide more details about the end-to-end encryption implementation?

Why This Matters:

  • The explicit follow-up triggers better semantic search
  • The AI retrieves more relevant sections from your documents
  • You get more accurate, detailed answers

Note: The AI still understands conversational references - this is specifically about optimizing document retrieval for RAG-based responses.

Breaking Down Complex Requests​

Complex, multi-part tasks are better handled as a series of focused questions rather than one large request.

When to Break Down:

  • Tasks requiring multiple steps or analyses
  • Questions combining several different topics
  • Requests involving multiple documents or data sources
  • Tasks requiring both research and creation
Example: Breaking Down a Complex Task

❌ Too Complex - Single Request:

"Analyze our Q3 financial data, compare it to Q2 and last year's Q3,
identify trends in customer acquisition costs, create projections for Q4,
and draft an executive summary with recommendations for budget allocation
across marketing channels, formatted as a PowerPoint presentation."

✅ Better Approach - Sequential Requests:

1. "Analyze the Q3 financial data and summarize key metrics"
2. "Compare these Q3 results with Q2 and last year's Q3"
3. "What trends do you see in customer acquisition costs over these periods?"
4. "Based on these trends, what are realistic projections for Q4?"
5. "Draft an executive summary highlighting the main findings"
6. "What budget allocation across marketing channels would you recommend?"
7. "Structure this information for a PowerPoint presentation with slide titles"

Benefits:

  • Each response can be reviewed and verified before proceeding
  • You can course-correct if the AI misunderstands a step
  • The AI can focus deeply on each component
  • Results are more accurate and thorough
  • You maintain better control over the output
Verification for Critical Information

For important decisions or high-value information, always verify sources and claims:

  • Check the sources provided with web search results
  • Review the document sections the AI references
  • Validate numerical data against original sources
  • For agents with database connections, verify query results

Learn more about transparency and explainability features in Agent Configuration.

Working with Conversation History and Context​

Understanding how the AI uses conversation context helps you structure your interactions for optimal results.

Context Window Limits:

While the AI effectively maintains conversation context for typical usage, extremely long conversations may hit context window limits:

  • Reasoning Model: Large context window - handles extended conversations well
  • Fast Model: Smaller context window - rarely an issue for normal usage

Signs You've Reached Context Limits:

  • The AI forgets information from early in the conversation
  • Responses become inconsistent with earlier statements
  • The AI seems confused about what was previously discussed
  • Answer quality noticeably degrades

What to Do:

  1. Summarize key points from the conversation so far
  2. Start a fresh conversation with the summary as context
  3. Reference the most important information rather than entire conversation history

When to Start Fresh​

Best Practice: One Conversation, One Topic

Keeping conversations focused on a single topic or project helps the AI provide more relevant, coherent responses.

Start a New Conversation When:

  • Switching to a completely different topic or project
  • Beginning a new task that doesn't relate to the current discussion
  • The current conversation has become very long
  • You need a fresh perspective without prior context influencing answers
  • You want to compare different approaches to the same problem

Why This Matters:

  • The AI focuses on relevant context without distraction from unrelated information
  • Conversation history stays organized and easy to find later
  • You avoid context pollution from mixing multiple topics
  • The AI provides more targeted, on-point responses

Example Structure:

  • Conversation 1: "Q3 Financial Analysis - Customer Acquisition"
  • Conversation 2: "Q3 Financial Analysis - Revenue Breakdown"
  • Conversation 3: "Product Roadmap Planning - Feature Prioritization"

Rather than one long conversation covering all three topics.

Web Search Best Practices​

Web search allows the AI to access current information from the internet, but understanding its capabilities and limitations helps you use it effectively.

Quick Guide:

  • Web search performs 2-3 targeted searches per query
  • Works like manual browsing - no special access to paywalled content
  • Best for specific, focused questions about current information
Good Uses vs. Not Suitable For

Good Uses for Web Search:

✅ Looking up current news or recent developments

✅ Finding product specifications or pricing

✅ Checking the latest version of software or tools

✅ Researching recent industry trends or reports

✅ Verifying up-to-date facts or statistics

✅ Finding complementary information to augment uploaded documents

Not Suitable for Web Search:

❌ Comprehensive research requiring dozens of sources

❌ Questions like "What is the current legislation on X for country Y?" (requires extensive research)

❌ Deep academic or scientific literature reviews

❌ Tasks requiring access to specialized databases

❌ Comparative analysis of multiple complex topics

Example: Breaking Down Broad Research Requests

❌ Too Broad:

"Research the entire competitive landscape for project management software,
including features, pricing, market share, recent acquisitions, and customer
sentiment across all major players."

✅ Focused Questions:

1. "What are the top 5 project management software platforms by market share in 2024?"
2. "What are the current pricing tiers for Asana and Monday.com?"
3. "What are recent customer reviews saying about Jira's new features?"
Hallucination Risk

If you ask a question that requires extensive web research, the AI may attempt to answer using its pretrained knowledge instead of admitting it needs more searches. This increases the risk of hallucination (confidently stating incorrect information).

Prevention: Explicitly instruct the AI to only use information from the web search results, or break your question into smaller, searchable pieces.

Potential and Limitations​

Key Point: The AI navigates websites like a human user - it encounters the same limitations you would.

Website Access Limitations and Solutions

May Encounter:

  • Websites with poorly structured content that's difficult to extract
  • JavaScript-heavy sites that don't render properly for automated access
  • Paywalls or login requirements that block content
  • CAPTCHAs or bot-detection mechanisms
  • Rate limiting from websites preventing access
  • Geo-restricted content not available in the server's location

Works Best With:

  • Static content websites with clear HTML structure
  • News sites and blogs with accessible articles
  • Public documentation and knowledge bases
  • Government and educational resources
  • Well-structured corporate websites

If Web Search Fails:

  • Try rephrasing your question to target different sources
  • Specify alternative sources the AI should try
  • Consider uploading documents directly instead
  • Check if the information is available through other means
  • Report technical issues to your administrator via the support ticketing system

Directing Web Sources​

You can guide the AI toward specific sources or away from unreliable ones.

Specifying Sources:

"Search for information about climate change, but only use sources from
academic institutions and government research agencies."
"Find recent reviews of the iPhone 15, focusing on technology news sites
like TechCrunch, The Verge, and Ars Technica."
"Look up the latest React.js best practices, but only reference the official
React documentation and well-known developer resources."

Avoiding Sources:

"Research market trends but avoid opinion pieces and blogs - focus on
analyst reports and industry publications."

Why This Helps:

  • Improves reliability and credibility of information
  • Targets authoritative sources for your domain
  • Avoids contradictory information from questionable sources
  • Ensures consistency with your organization's standards
If You Encounter Issues:

If web search consistently fails or behaves unexpectedly, it may indicate configuration issues or tool problems. Report issues to your administrators through the support ticketing system, you can find more information here: Support Ticketing Tutorial.

File Upload and Management​

Understanding how the AI processes and uses files helps you get better results from document analysis.

How Files Are Processed​

Overview: Galene.AI uses multimodal analysis to understand both text content and visual layout.

Technical Details: Document Processing

Galene.AI uses a sophisticated approach to extract information from documents:

  1. Text Chunking: Documents are divided into meaningful segments
  2. Summary Extraction: Summaries provide context for each chunk
  3. Multimodal Embeddings: Both text and visual information are analyzed
  4. Visual Layout Analysis: When layout or visual content is important, the AI also receives page images to analyze graphically

Benefits:

  • Comprehensive information extraction from both text and layout
  • No OCR needed
  • Efficient processing times
  • Understanding of complex document structures
  • Flexible responses based on content and context

Works Well For:

  • PDFs with embedded text and tables
  • Word documents with complex formatting
  • Presentations with text and diagrams
  • Documents where layout conveys meaning
  • Reports with charts, tables, and callouts

Spreadsheets and Structured Data​

Important Limitation:

Spreadsheets (Excel, CSV files) are processed as documents for semantic search. The AI:

  • ✅ Can read and understand tabular data
  • ✅ Can answer questions about specific content
  • ❌ Cannot perform advanced filtering or complex calculations
  • ❌ Cannot execute spreadsheet functions or formulas
  • ❌ Should not be used as a database querying tool
For Advanced Data Analysis

For querying numerical data, complex filtering, and database operations, use Database Connectors configured by your administrators. Database connections enable natural language queries against structured data sources for sophisticated analysis.

When the AI Can't Find Information​

If the AI says it cannot find information that should be in uploaded files, try these steps:

Troubleshooting Steps

1. Verify the File Is Uploaded or in the Knowledge Base

If the file should have been uploaded in the current chat:

  • Check the File Management panel in the right sidebar
  • Ensure the file status shows "Processed"
  • Confirm the file checkbox is enabled (active)

If it should be in the Agent knowledge base instead:

  • Open the Agent configuration
  • Check that the file has correctly been uploaded and processed

2. Check Your Query

  • Make the question more specific and detailed
  • Include relevant keywords from the document
  • Rephrase using different terminology

3. Break Down the Question

  • Ask about one aspect at a time
  • Start with broader questions, then get specific

Example of Refinement:

❌ "What does it say about the requirements?"

↓

🔄 "What are the technical requirements mentioned in the document?"

↓

✅ "What are the database and server requirements for the deployment process described in section 3?"

Quick Fix

Often, the issue can be solved by rephrasing the question. More specific queries trigger better semantic search results.

File Organization Best Practices​

Choose Descriptive Filenames:

Good file naming helps you and the AI understand content context:

✅ Clear and Specific:

  • Q3_2024_Financial_Report_Final.pdf
  • Product_Spec_v2.1_Updated_Dec2024.docx
  • Customer_Survey_Results_Enterprise_Segment.xlsx

❌ Vague or Generic:

  • Document1.pdf
  • Report.docx
  • Data.xlsx
  • Final_FINAL_v2_really_final.pdf

Why This Matters:

  • Easier to identify the right file when managing multiple documents
  • Helps you understand document content at a glance
  • Version tracking becomes clear and manageable
  • Particularly important when dealing with similar content types

Managing Files Within a Conversation:

  • Avoid information overload: Don't upload an excessive number of similar documents that might confuse the AI
  • Focus conversations: Keep one conversation focused on one topic - avoid mixing files from unrelated projects
  • Use File Management: Actively use checkboxes in the File Management sidebar to enable/disable files based on your current question
  • Be specific with comparisons: When comparing documents, ask specific questions rather than broad comparisons across many files

Exporting AI Responses:

  • The chat interface cannot create downloadable files of any format
  • To save responses, copy the text from the chat
  • For formatted content, copy from markdown code blocks
  • Consider pasting into your preferred document editor for further formatting

Response Optimization​

Iterative Refinement​

The AI's responses can be progressively improved through follow-up questions and clarifications.

Iterative Approach:

  1. Initial Request → Start with your core question
  2. Review Response → Evaluate the answer
  3. Refine → Ask for adjustments or expansions
  4. Repeat → Continue until satisfied
Example: Step-by-Step Refinement Process

Example Iteration:

You: "Explain the benefits of microservices architecture"
AI: [Provides general explanation]
You: "Make this more technical, suitable for senior engineers"
AI: [Provides detailed technical explanation]
You: "Add specific examples from e-commerce platforms"
AI: [Adds relevant examples]
You: "Format this as bullet points with code snippets where relevant"
AI: [Reformats with code examples]

Format Specification Examples:

  • "Provide this as a bulleted list"
  • "Create a comparison table with 3 columns"
  • "Write this as an executive summary (max three paragraphs)"
  • "Format as a step-by-step tutorial with numbered steps"
  • "Present findings as a formal business report with sections"
When to Report Issues:

If the AI consistently fails to meet your needs, produces errors, or behaves unexpectedly, report the issue to your administrators through the support ticketing system so they can investigate and improve the system. See Support Ticketing for more details.

Validation and Verification​

For critical information and important decisions, always verify the AI's responses.

Verification Strategies by Source Type

For Web Search Results:

  • Review the sources provided at the end of responses
  • Click through to original sources to verify context
  • Check the publication date and author credibility
  • Cross-reference with multiple sources when possible

For Document Analysis:

  • Ask the AI to cite specific sections from the document
  • Manually verify key quotes and data points
  • Check that interpretations align with document intent
  • Review the actual document sections referenced

For Database Queries:

  • Review the query logic the agent used
  • Verify results against known data patterns
  • Cross-check critical numbers with original databases
  • Understand the query limitations and scope

Explainability Features:

Galene.AI provides transparency features to help you verify information:

  • Source Citations: See which documents or web pages information came from
  • Query Transparency: For database agents, understand what queries were executed
  • Document References: View which sections of uploaded files were used

Learn more about these features in Agent Configuration.

Critical Information Verification

Never rely solely on AI-generated information for:

  • Legal or compliance decisions
  • Financial transactions or reporting
  • Medical or safety-critical decisions
  • Mission-critical business decisions

Always verify through appropriate channels and human expertise.

Data Privacy and Security Practices​

Understanding Data Handling​

Galene.AI takes data security seriously, but users should still follow best practices when sharing information with each other.

Agent Sharing Considerations:

  • Before sharing an agent: Verify that recipients have appropriate authorizations for any uploaded content and data sources
  • Sensitive company data: Avoid sharing agents or conversations containing confidential information unless recipients have proper clearance
  • Role-based access: If an agent has database or knowledge base connectors, sharing will be automatically prevented if the recipient doesn't have adequate access roles

Connector Permissions:

When agents are connected to enterprise data sources (databases, knowledge bases, MCP servers), access is controlled by role-based permissions. This ensures that sensitive data remains protected even when agents are shared.

Learn more about connectors and permissions in Connecting to Enterprise Data.

Sensitive Information Guidelines​

Galene.AI Security Model:

Galene.AI is designed with security and privacy as core principles:

Your Data Security:

  • On-premise deployment: Your data stays within your infrastructure
  • Encrypted storage: All conversations and data are encrypted
  • Private conversations: Your conversations are completely private
  • Audit access only: Conversation data can only be accessed for legal and audit purposes

Your Rights:

  • Intellectual property: You retain full IP rights to AI-generated outputs
  • No consumption costs: Use the platform freely without per-query charges
  • No training on your data: Models are never trained on your conversation data
  • Complete privacy: Write and upload whatever you need without concerns

Best Practices:

While Galene.AI provides enterprise-grade security, always:

  • Follow company policies: Adhere to your organization's data handling guidelines
  • Apply need-to-know principles: Only share information with those who require it
  • Document usage: Keep records of how AI is used for critical decisions
  • Report concerns: Alert administrators if you notice unusual behavior or potential security issues

Freedom to Use:

With Galene.AI's security model, you can:

  • Upload any business documents without privacy concerns
  • Discuss confidential projects and strategies
  • Analyze sensitive financial or operational data
  • Work with proprietary information

All within your organization's data governance framework.

Getting Help​

If you encounter issues or have questions not covered in this guide:

  • Technical Issues: Report problems through your organization's support ticketing system
  • Feature Requests: Work with your administrators to request new capabilities
  • Training: Consult with your team's Galene.AI administrators for additional training
  • Documentation: Refer to other sections of this documentation for specific features

Next Steps​

Now that you understand best practices for using Galene.AI effectively: