Build an All-Source Knowledge Assistant with Claude, RAG, Perplexity, and Drive

Description

📜 Detailed n8n Workflow Description

Main Flow

The workflow operates through a three-step process that handles incoming chat messages with intelligent tool orchestration:

Message Trigger: The When chat message received node triggers whenever a user message arrives and passes it directly to the Knowledge Agent for processing.

Agent Orchestration: The Knowledge Agent serves as the central orchestrator, registering a comprehensive toolkit of capabilities:
LLM Processing: Uses Anthropic Chat Model with the claude-sonnet-4-20250514 model to craft final responses
Memory Management: Implements Postgres Chat Memory to save and recall conversation context across sessions
Reasoning Engine: Incorporates a Think tool to force internal chain-of-thought processing before taking any action
Semantic Search: Leverages General knowledge vector store with OpenAI embeddings (1536-dimensional) and Cohere reranking for intelligent content retrieval
Structured Queries: Provides structured data Postgres tool for executing queries on relational database tables
Drive Integration: Includes search about any doc in google drive functionality to locate specific file IDs
File Processing: Connects to Read File From GDrive sub-workflow for fetching and processing various file formats
External Intelligence: Offers Message a model in Perplexity for accessing up-to-the-minute web information when internal knowledge proves insufficient

Response Generation: After invoking the Think process, the agent intelligently selects appropriate tools based on the query, integrates results from multiple sources, and returns a comprehensive Markdown-formatted answer to the user.

Persistent Context Management

The workflow maintains conversation continuity through Postgres Chat Memory, which automatically logs every user-agent exchange. This ensures long-term context retention without requiring manual intervention, allowing for sophisticated multi-turn conversations that build upon previous interactions.

Semantic Retrieval Pipeline

The semantic search system operates through a sophisticated two-stage process:

Embedding Generation**: Embeddings OpenAI converts textual content into high-dimensional vector representations
Relevance Reranking**: Reranker Cohere reorders search hits to prioritize the most contextually relevant results
Knowledge Integration**: Processed results feed into the General knowledge vector store, providing the agent with relevant internal knowledge snippets for enhanced response accuracy

Google Drive File Processing

The file reading capability handles multiple formats through a structured sub-workflow:

Workflow Initiation: The agent calls Read File From GDrive with the selected fileId parameter
Sub-workflow Activation: When Executed by Another Workflow node activates the dedicated file processing sub-workflow
Operation Validation: Operation node confirms the request type is readFile
File Retrieval: Download File1 node retrieves the binary file data from Google Drive
Format-Specific Processing: FileType node branches processing based on MIME type:
PDF Files: Route through Extract from PDF → Get PDF Response to extract plain text content
CSV Files: Process via Extract from CSV → Get CSV Response to obtain comma-delimited text data
Image Files: Analyze using Analyse Image with GPT-4o-mini to generate visual descriptions
Audio/Video Files: Transcribe using Transcribe Audio with Whisper for text transcript generation
Content Integration: The extracted text content returns to Knowledge Agent, which seamlessly weaves it into the final response

External Search Capability

When internal knowledge sources prove insufficient, the workflow can access current public information through Message a model in Perplexity, ensuring responses remain accurate and up-to-date with the latest available information.

Design Highlights

The workflow architecture incorporates several key design principles that enhance reliability and reusability:

Forced Reasoning**: The mandatory Think step significantly reduces hallucinations and prevents tool misuse by requiring deliberate consideration before action
Template Flexibility: The design is intentionally generic—organizations can replace **[your company] placeholders with their specific company name and integrate their own credentials for immediate deployment
Documentation Integration**: Sticky notes throughout the canvas serve as inline documentation for workflow creators and maintainers, providing context without affecting runtime performance

System Benefits

With this comprehensive architecture, the assistant delivers powerful capabilities including long-term memory retention, semantic knowledge retrieval, multi-format file processing, and contextually rich responses tailored specifically for users at [your company]. The system balances sophisticated AI capabilities with practical business requirements, creating a robust foundation for enterprise-grade conversational AI deployment.