Context Q&A Agent

​

Overview

​

Initialization

# agent/agent_factory.py:264-290
context_qna_agent = Agent(
    id="context-qna-agent",
    name="Chat Context Q&A",
    role="Answering questions about users, topics, and past conversations based on extensive chat history",
    model=OpenAILike(
        id=CONTEXT_AGENT_MODEL,
        max_tokens=8000,
        temperature=0.3,
        base_url=PROVIDER,
        api_key=CUSTOM_PROVIDER_API_KEY,
    ),
    tools=[HistoryTools(), BioTools(client=client)],
    add_datetime_to_context=True,
    timezone_identifier="Asia/Kolkata",
    instructions="""You specialize in answering questions about the chat history, users, and topics discussed.

You have access to `read_chat_history`. Call this tool to get the conversation history before answering questions.
IMPORTANT: always fetch a minimum of 5000 messages on first try.
Use the history to:
- Answer "who said what" questions
- Summarize discussions on specific topics
- Track when topics were last mentioned
- Identify user opinions and statements
- Provide context about past conversations

Be precise with timestamps and attribute statements accurately to users."""
)

​

Model Configuration

model=OpenAILike(
    id=CONTEXT_AGENT_MODEL,  # From environment
    max_tokens=8000,
    temperature=0.3,  # Lower temp for factual recall
    base_url=PROVIDER,
    api_key=CUSTOM_PROVIDER_API_KEY,
)

​

Model Requirements

• Long context window: Must handle thousands of messages
• Low temperature: 0.3 for accurate fact retrieval
• High token limit: 8000 for comprehensive responses

CONTEXT_AGENT_MODEL="gemini-2.0-flash-exp"  # Or other long-context model
CONTEXT_AGENT_MAX_MESSAGES=5000  # Default minimum fetch

​

Tools

​

1. HistoryTools

tools=[HistoryTools(), BioTools(client=client)]

IMPORTANT: always fetch a minimum of 5000 messages on first try.

1. Call read_chat_history before answering
2. Fetch at least 5000 messages initially
3. Fetch more if needed for complete context

​

2. BioTools

BioTools(client=client)

• User profiles
• Stored preferences
• User metadata

​

Use Cases

Long-context chat history / thread analysis / who-said-what → delegate to context-qna-agent (requires Channel ID).

​

1. Message Attribution

• “Who said they were working on the API?”
• “Did anyone mention the deadline?”
• “What did Alice say about the project?”

1. Fetches history with read_chat_history
2. Searches for relevant messages
3. Attributes statements to specific users
4. Includes timestamps

​

2. Topic Summaries

• “Summarize the discussion about deployment”
• “What have we decided about the new feature?”
• “Recap the meeting from yesterday”

1. Retrieves extensive message history
2. Identifies topic-relevant messages
3. Chronologically summarizes
4. Notes key decisions and opinions

​

3. Timeline Tracking

• “When was the bug first mentioned?”
• “How long ago did we discuss pricing?”
• “What was the last message about the release?”

1. Scans timestamps in history
2. Identifies first/last mentions
3. Provides precise timing
4. Shows evolution of discussion

​

4. Opinion Mining

• “What’s Bob’s opinion on the redesign?”
• “Who supports the new approach?”
• “What concerns were raised?”

1. Finds user-specific statements
2. Extracts opinions and sentiments
3. Provides context for each opinion
4. Notes changes over time

​

Instructions

You specialize in answering questions about the chat history, users, and topics discussed.

You have access to `read_chat_history`. Call this tool to get the conversation history before answering questions.
IMPORTANT: always fetch a minimum of 5000 messages on first try.

Use the history to:
- Answer "who said what" questions
- Summarize discussions on specific topics
- Track when topics were last mentioned
- Identify user opinions and statements
- Provide context about past conversations

Be precise with timestamps and attribute statements accurately to users.

​

Context Awareness

add_datetime_to_context=True,
timezone_identifier="Asia/Kolkata",

​

Temporal Context

The current date/time is provided at the start of the context in IST. Each message has a timestamp like [2h ago], [1d ago], or [Dec 15, 14:30] - all times are in IST.

​

Past vs Present

When someone says “I’m working on X” in a message from 2 hours ago, they were working on it THEN, not necessarily now.

• Uses phrases like “Earlier you mentioned…”
• Distinguishes past statements from current reality
• Calculates relative times (“yesterday”, “last week”)

​

Reply Context

If a user is replying to a specific message, you will see a [REPLY CONTEXT] block before their message.

​

Channel ID Requirement

delegate to context-qna-agent (requires Channel ID)

• Scope history retrieval
• Access correct conversation thread
• Maintain context boundaries

​

Performance Characteristics

• Speed: Moderate (history fetch adds latency)
• Accuracy: High (low temperature, precise retrieval)
• Context limit: Determined by CONTEXT_AGENT_MODEL
• Minimum fetch: 5000 messages per query

​

Best Practices

​

1. Always Fetch History First

# CORRECT workflow:
1. User asks: "What did I say about the API?"
2. Agent calls: read_chat_history(limit=5000)
3. Agent analyzes history
4. Agent responds with attributed quote

# INCORRECT:
1. User asks: "What did I say about the API?"
2. Agent guesses or says "I don't know"

​

2. Fetch Sufficient Messages

# Minimum first fetch
read_chat_history(limit=5000)

# Fetch more if needed
if topic_not_found and messages_available:
    read_chat_history(limit=10000)

​

3. Precise Attribution

GOOD: "Alice said 'the API is ready' on Dec 15 at 14:30 IST"
BAD: "Someone mentioned the API might be ready"

​

4. Temporal Precision

GOOD: "The bug was first mentioned 3 days ago by Bob"
BAD: "The bug was mentioned recently"

​

Configuration

# Long-context model for history analysis
CONTEXT_AGENT_MODEL="gemini-2.0-flash-exp"

# Maximum messages to consider
CONTEXT_AGENT_MAX_MESSAGES=5000

# Provider endpoint
PROVIDER=https://your-openai-compatible-endpoint
CUSTOM_PROVIDER_API_KEY=your_api_key

​

Delegation Strategy

​

When to Use Context Q&A Agent

• Questions about past conversations
• “Who said” queries
• Topic summaries
• Timeline questions
• Opinion tracking

• Current web information (use Perplexity Agent)
• Code execution (use Code Agent)
• Math (use Team Leader)
• Real-time data (use Perplexity Agent)

​

Next Steps

• History Tools Integration
• Memory Management
• System Prompts