Skip to content

LLM Processing

Model Configuration

The system uses Anthropic's Claude as the primary AI engine, with per-agent model selection possible.

Setting Value
Model claude-sonnet-4-20250514
Temperature 0.1 (for consistent responses)
Streaming Enabled
Framework LangGraph + @langchain/langgraph-supervisor

Message Processing Flow

sequenceDiagram
    participant User
    participant MessagesService
    participant Supervisor
    participant Subagent
    participant MCP as MCP Tools

    User->>MessagesService: Send message
    MessagesService->>MessagesService: Build context (history, memory)
    MessagesService->>MessagesService: Load MCP tools, partition by @agents
    MessagesService->>Supervisor: Stream (messages + tools + updates)
    Supervisor->>Supervisor: Decide which agent handles request
    Supervisor->>Subagent: transfer_to_intelligence_agent
    Subagent->>MCP: Call tools (SearchCompanies, FindRegions...)
    MCP-->>Subagent: Tool results
    Subagent-->>Supervisor: Response + tool history
    Supervisor-->>User: Stream response text
    MessagesService->>MessagesService: Save conversation, token usage

Cross-Domain Queries

For requests that span multiple domains (e.g., "Find SaaS companies and create a list"):

  1. Supervisor calls intelligence_agent first → gets search results
  2. Supervisor calls lists_agent with the results → creates the list
  3. Supervisor composes the final response

Streaming Architecture

Three stream modes run simultaneously:

Mode What it captures When it fires
messages LLM token-by-token text output Real-time as LLM generates
tools Tool lifecycle events (on_tool_start, on_tool_end) Real-time as tools execute
updates Node-level state updates (token counts, conversation tracking) After each graph node completes

Handoff Tool Filtering

The supervisor uses internal transfer_to_* and transfer_back_to_* tools for routing. These are filtered from:

  • Streaming events — isHandoffToolName() in stream-handler.util.ts
  • Conversation history — Filtered in DatabaseConversationHistoryAdapter to prevent Anthropic's tool_use/tool_result pairing errors on subsequent turns

System Prompts

Supervisor Prompt

Pure routing logic — never answers questions directly:

  • Lists available agents and their capabilities
  • Routing rules for common query types
  • Dynamically includes customer_tools_agent when customer servers are connected
  • Memory block and conversation history appended at runtime

Domain Prompts

Each subagent has a focused prompt in src/connectors/prompts/subagent-prompts/:

Prompt Content
intelligence.ts Entity search rules, filter syntax, financial data formatting
lists.ts 4-step list creation workflow, campaign management
analytics.ts Dashboard generation, ClickHouse queries, scoring
workflows.ts Workflow creation sequence, trigger types
data-ops.ts Import templates, source management, matching review
reports.ts Report builder workflow, template instructions
data-science.ts Pipeline operations
shared-formatting.ts Citation format, number formatting, work page links (shared across agents)

Debug Modes

Mode Description
[debug] Full MCP Request/Response debug with tool call details
[info] Basic tool call notifications
Default Standard processing without debug output

File Processing

Supported File Types

Text: PDF, TXT, HTML Structured Data: CSV, Excel (XLSX/XLS), JSON Media: Images (JPG, PNG, GIF, WebP), Audio (MP3, WAV, M4A, OPUS with transcription) Documents: DOC/DOCX, PPT/PPTX

Document Export

The export_document tool (on the supervisor) generates downloadable PDF or DOCX files from text content.