AI Chat & Research
Master the AI chat interface to get the most value from your document analysis.
How AI Chat Works
BillionLens uses an advanced Multi-Query Retrieval-Augmented Generation (RAG) pipeline:
- Query Decomposition → Complex questions are broken into focused sub-queries
- Parallel Vector Search → Multiple searches execute against the document index
- Context Aggregation → Results are deduplicated and ranked by relevance
- Grounded Generation → AI synthesizes answers constrained to retrieved evidence
- Citation Verification → Every claim is verified against source documents
The Chat Interface
The chat interface has three columns:
| Column | Purpose |
|---|---|
| Left Sidebar | Conversation history - switch between conversations |
| Center | Chat messages - ask questions and see responses |
| Right Panel | Source citations - documents supporting the answer |
Starting a Conversation
- Navigate to your case
- Click Chat with AI
- Type your question and press Enter or click Send
- AI will stream the response in real-time
Managing Conversations
- New Conversation - Click "New Conversation" to start a fresh topic
- Switch Conversations - Click any conversation in the sidebar
- Rename - Click the conversation to select it, then rename in the input
- Delete - Click the trash icon to remove a conversation
Writing Effective Questions
✅ Good Questions
Specific and focused:
What emails did Sarah Johnson send about the merger between March and June 2024?
Request summaries:
Summarize the key terms of the Operating Agreement.
Ask about people:
What is John Smith's role in the company?
Seek connections:
What documents mention both Acme Corp and the investment agreement?
Financial questions:
What were the capital contributions made by each member?
❌ Less Effective Questions
Too broad:
Tell me about this case.
Better: "Summarize the main parties and their roles in the Operating Agreement."
Asking AI to make legal judgments:
Will we win this case?
Better: "What evidence in the documents supports breach of fiduciary duty?"
Outside document scope:
What is the current stock price of ABC Corp?
Better: Questions should be about your uploaded documents.
Using Metadata Filters
Narrow your search using metadata filters:
Available Filters
| Filter | Purpose |
|---|---|
| Custodian | Filter by document owner |
| File Type | Filter by format (PDF, DOCX, etc.) |
| Document Type | Filter by category (Email, Contract, etc.) |
| Confidentiality | Filter by classification level |
| Date Range | Filter by document dates |
Example: Searching Specific Custodian
- Click the filter button
- Select "Custodian: John Smith"
- Ask: "What emails mention the settlement?"
- Results will only come from John Smith's documents
Understanding Citations
Source Panel
Every AI response shows supporting sources in the right panel:
- File Name - Original document name
- Production Number - Bates/production number if available
- Custodian - Document owner
- Document Type - Category (Email, Contract, etc.)
- Date - Document date
- Download Link - Click to download the file
Citation Quality
BillionLens uses grounded generation - every claim is backed by source documents:
- ✅ Answers are verified against your documents
- ✅ No hallucination - only information from your files
- ✅ Citations allow you to verify every claim
- ✅ Confidence based on source quality
When AI Says "I Don't Know"
If the AI can't find relevant information:
- The documents may not contain the answer
- Try rephrasing your question
- Check if all relevant documents are uploaded and indexed
- Try removing metadata filters to broaden the search
Advanced Techniques
Follow-up Questions
Build on previous answers:
You mentioned three contracts. Can you provide more details about the first one?
Comparative Questions
How do the terms in Exhibit A differ from Exhibit B?
Timeline Questions
What events occurred in August 2024 according to the documents?
Evidence Gathering
What evidence supports the claim that John made capital contributions?
Best Practices
📝 Start Broad, Then Narrow
- First question: "Who are the key parties in this matter?"
- Follow-up: "Tell me more about Michael Chen's role."
- Specific: "What is Michael Chen's ownership percentage?"
🎯 One Topic Per Conversation
Do:
- Create separate conversations for distinct topics
- "Ownership & Contributions" conversation
- "Timeline of Events" conversation
Don't:
- Mix unrelated questions in one conversation
- Jump between completely different subjects
🔍 Verify Critical Information
Always verify AI responses for important facts:
- Check cited documents directly
- Cross-reference with your own review
- Use AI as research assistant, not final authority
💡 Types of Questions
Document Summary:
Summarize the main points of the Operating Agreement.
People Search:
Who are the founding members and what are their roles?
Financial Analysis:
What capital contributions were made and by whom?
Timeline Building:
Create a chronological list of key events in this matter.
Conversation History
Automatic Saving
All conversations are automatically saved:
- Questions and answers preserved
- Source citations stored
- History available across sessions
Organizing Conversations
Give conversations descriptive names:
- ✅ "Ownership Analysis"
- ✅ "Financial Transactions August 2024"
- ✅ "Key Personnel Roles"
- ❌ "New Conversation 1"
Limitations
What AI Can Do
✅ Search thousands of documents instantly ✅ Identify patterns and connections ✅ Summarize complex documents ✅ Extract relevant quotes with citations ✅ Answer questions about document content
What AI Cannot Do
❌ Provide legal advice or strategy ❌ Make subjective judgments ❌ Access information not in your documents ❌ Predict case outcomes ❌ Replace attorney judgment
Accuracy
- AI is highly accurate but not perfect
- Always verify critical facts against source documents
- Check citations for important claims
- Use as starting point for deeper analysis
Privacy & Security
Data Handling
Your documents are processed with enterprise-grade security:
- Files stored in encrypted, geographically redundant cloud storage
- AI processing occurs in isolated sessions with no data retention
- Document content is never used for model training
- All transmissions encrypted with TLS 1.3
Best Practices
- Verify critical information against source documents
- Use professional judgment about sensitive queries
- Maintain standard privilege review protocols
Next: Learn about Entity Extraction for automated analysis.