Skip to content

agent-engine.md

Latest commit

 

History

History
84 lines (65 loc) · 2.51 KB

agent-engine.md

File metadata and controls

84 lines (65 loc) · 2.51 KB

Agent Engine

Overview

The agent engine is the core AI workflow built on LangGraph. It processes emails through a deterministic pipeline of nodes, each powered by Google Gemini via LangChain.

Node Pipeline

1. Classification Node (classify)

  • Calls GeminiClient.classify_intent() via LangChain
  • Returns: intent category + confidence score
  • Short-circuits spam emails with high confidence

2. Context Retrieval Node (retrieve)

  • Embeds email text using Gemini embeddings (via LangChain)
  • Searches pgvector for similar past emails
  • Fetches CRM contact data and calendar events
  • Returns: ranked list of context strings

3. Decision Node (decide)

  • Analyzes classification + context
  • Selects tools to invoke from available registry
  • Returns: list of tools with parameters

4. Tool Execution Node (execute_tools)

  • Iterates selected tools via ToolManager
  • Each call logged to tool_executions table
  • Failures isolated — one tool failing doesn't block others
  • Returns: dict of tool results

5. Generation Node (generate)

  • Generates email response using Gemini via LangChain
  • Incorporates context + tool results
  • Returns: draft response text + confidence

6. Review Node (review)

  • Decides: auto-approve vs human review
  • Threshold: confidence ≥ 0.8 → auto-approve
  • Complaints always require human review
  • Returns: requires_approval flag

7. Dispatch Node (dispatch)

  • Sends approved response via Gmail API
  • Updates email status to "sent"

LangChain + Gemini Integration

All LLM calls use langchain-google-genai:

from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(
    model="gemini-2.0-flash",
    google_api_key=settings.GEMINI_API_KEY,
)

Embeddings use:

from langchain_google_genai import GoogleGenerativeAIEmbeddings

embeddings = GoogleGenerativeAIEmbeddings(
    model="models/embedding-001",
    google_api_key=settings.GEMINI_API_KEY,
)

Tool Registry

Tool Integration Description
send_email Gmail Send an email response
create_draft Gmail Create a draft for review
check_calendar Calendar Check availability
create_event Calendar Schedule a meeting
get_contact CRM Fetch contact info
update_contact CRM Update contact data

Observability

Every execution generates:

  • AgentLog entries for each node (with full state I/O)
  • ToolExecution entries for each tool call
  • A trace_id to group all entries for one run