Skip to content

LLM Processing

Model Configuration

The assistant uses Anthropic's Claude, selected per purpose. The customer chat path (AGENT_CHAT) runs a single agent loop.

Setting Value
Model claude-sonnet-4-5
Max output tokens 4096
Temperature 0.1 (for consistent responses)
Streaming Enabled
Framework LangChain createAgent (single agent — no supervisor)

"Max output tokens" is the Anthropic max_tokens cap — the ceiling on the length of a single generated reply, not a context-window or input budget. The model and this cap are not hardcoded in the chat path: they come from the active provider's entry in LLM_PROVIDER_CONFIG (ACTIVE_LLM_PROVIDER, currently anthropic → claude-sonnet-4-5 / 4096). Switching the active provider changes both, and either can be overridden per call. Other purposes pin their own values (e.g. memory/summarization on claude-haiku-4-5) — see llm-client.service.ts (PURPOSE_DEFAULTS) and llm-provider.config.ts.

Message Processing Flow

One agent handles the whole turn — it plans, calls tools, and answers in a single flat loop. There is no routing supervisor and no subagent handoffs.

sequenceDiagram
    participant User
    participant MessagesService
    participant Agent as Single Agent
    participant CLI as cred-platform CLI
    participant MCP as Residual MCP tools

    User->>MessagesService: Send message
    MessagesService->>MessagesService: Build context (history, memory)
    MessagesService->>Agent: Discover tools, select visible set, compose skills
    Agent->>CLI: e.g. lists create / company resolve / query
    CLI-->>Agent: Result
    Agent->>MCP: e.g. analytics / workflows / data-science
    MCP-->>Agent: Tool results
    Agent-->>User: Stream response text
    MessagesService->>MessagesService: Save conversation, token usage

Multi-step Queries

For a request like "Find SaaS companies and create a list", the single agent does it in one loop: it calls the CLI to search, then calls the CLI again to build the list from those results, then composes the reply — no cross-agent handoff, because there is only one agent.

Tool Surface & Selection

Each turn the connector assembles the visible tool set (selectVisibleTools):

  • cred_platform_cli — the primary data edge (lists, company/person resolve
  • read, filtering/screening, chart authoring, view/report assembly, schema introspection, raw GraphQL). Always included.
  • Residual MCP tools — analytics, workflows, reports, data-science, and customer-connected servers the CLI does not cover. Loaded every turn, then gated by the per-request source + vendor filters.
  • export_document — always included.
  • Built-in web search — appended by the caller.

The CLI-superseded MCP tools (entity search, list build, chart/view authoring) are dropped so the loop has one data path. See Architecture → Tool surface for the full supersede/exclude lists.

Streaming Architecture

Because there is no supervisor node, the stream is flat: every tool_use / tool_result is top-level, which lets the deterministic rollup fire off a tool result directly.

Mode What it captures When it fires
messages LLM token-by-token text output Real-time as the 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 (transfer_to_* / transfer_back_to_*) is a supervisor-only concern and does not apply to the single-agent path — there are no handoff tools in the stream.

System Prompt & Skills

The system prompt is assembled per turn by buildAssistantSystemPrompt:

  • A single, simplified base prompt (no supervisor routing prompt, no per-domain subagent prompts).
  • A CLI help snapshot so the agent knows the CLI's command surface.
  • Composed skill instructions — the skill registry resolves the skills matching the turn and folds their instructions in. Today the only skill is reports (build mechanics + a per-turn report template). See Architecture → skill system.
  • On the web surface, an optional persona block (workspace/user memory + instructions).

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 generates downloadable PDF or DOCX files from text content. It is one of the tools available to the single agent.