Agentic RAG Portfolio Assistant
A production-grade Agentic RAG system with LangGraph agent, Docling extraction, OpenSearch hybrid search, Cohere reranking, Redis caching, and RAGAS evaluation.
System Architecture
Overview
A recruiter visits my portfolio, clicks the AI assistant, and asks questions about my projects, skills, and experience. The system retrieves relevant information from my documents and generates accurate, grounded answers in seconds.
Unlike linear RAG pipelines, this system uses a LangGraph agent that reasons about which tools to call. The agent can search all documents, focus on specific categories, or look up a single file — then decide if it needs to retrieve again before generating an answer.
System Components
Input Guardrails
Custom Regex + LLMPrompt injection detection, harmful content filter, length validation
Portfolio Classifier
Keyword + LLM FallbackBlocks general knowledge — only answers questions about Adeesha's work
LangGraph Agent
StateGraph + 5 ToolsLLM decides which tools to call based on query intent. Max 2 retrieval rounds.
Hybrid Retrieval
OpenSearch (BM25 + k-NN)BM25 keyword search + k-NN dense vector search with RRF fusion
Reranking
Cohere rerank-v3.5Top 10 results narrowed to top 3 most relevant
Generation
Groq gpt-oss-120b + Gemini FallbackPrimary LLM with automatic fallback to Gemini 2.5 Flash on rate limits
Caching
Upstash RedisLLM response cache (1hr TTL) + conversation memory (24hr TTL)
Observability
LangfuseFull execution tracing, prompt versioning, grading scores
Agent Tools
| Tool | Description |
|---|---|
| search_all | Search all documents — default for most queries |
| search_projects | Focus on project documents only |
| search_skills | Focus on skills/resume documents |
| search_source | Search a specific file by name |
| list_documents | Show what's available in the index |
Key Features
Agentic RAG
LangGraph StateGraph with tool routing and multi-step retrieval
Parent-Child Chunking
2000 char parents (LLM context) + 500 char children (search precision)
Query Expansion
Synonym-based expansion for better BM25 recall
API Key Auth
X-API-Key header on all endpoints (health check open)
Rate Limiting
5 requests/min per IP
Gemini Fallback
Groq primary, Gemini 2.5 Flash on rate limit/connection errors
Tenacity Retry
3 attempts, exponential backoff on all external APIs
Tool Result Caching
In-memory 5min TTL avoids redundant search + rerank calls
Chunk Deduplication
Delete existing chunks before re-indexing
Structured Logging
logging module with per-module loggers (rag.graph, rag.tools, etc.)
API Endpoints
/Health check (OpenSearch, MinIO, Redis)/queryAsk a question (body: {question, source_filter?})/upload?folder=Upload file to MinIO (max 20MB)/files?folder=List files (optional folder filter)/files/{folder}/{filename}Delete file + reindex/reindexFull reindex from MinIO/clear-memoryClear conversation history/clear-cacheClear LLM response cacheEvaluation Metrics (RAGAS)
Faithfulness
Is the answer grounded in context (no hallucination)?
Answer Relevancy
Does the answer address the question?
Context Precision
Are retrieved docs focused and relevant?
Context Recall
Did we retrieve all needed information?
Factual Correctness
Is answer factually accurate vs ground truth?