researchAgent

ResearchAgent was inspired by a simple problem: asking an AI to research a topic often produces confident answers that are difficult to verify. We wanted to build a system where evidence, not the language model, decides what can be trusted. ResearchAgent takes a research question through a multi-agent workflow where agents retrieve relevant evidence, reason over multiple papers, use a knowledge graph to connect findings, and then independently verify and review the resulting claims. Every accepted finding is tied back to concrete evidence in the original PDFs with page, section, paragraph, and provenance information, while unsupported claims are rejected rather than silently completed. We built the system with Python, LangGraph, local Ollama models, Groq's GPT-OSS 120B, semantic and lexical retrieval, embeddings, a provenance-aware knowledge graph, and a zero-trust validation layer. Building it taught us that reliable agentic AI is less about adding more agents and more about controlling what agents are allowed to claim. One of our biggest challenges was handling failure honestly: we encountered hallucinated citations, retrieval failures, budget exhaustion, model rate limits, graph persistence issues, and reasoning models that could generate plausible claims that our verifier rejected. Instead of hiding these failures, we designed the system to surface them and recover where possible. The result is an evidence-first research agent that doesn't just generate an answer—it investigates, challenges its own conclusions, and shows exactly why a conclusion should or should not be trusted.

Built With

• agentic-ai
• automated-research
• bm25
• embeddings
• evidence-grounding
• fastapi
• gpt-oss-120b
• groq
• information-retrieval
• knowledge-graph
• langchain
• langgraph
• llms
• multi-agent-systems
• neo4j
• ollama
• provenance-tracking
• pymupdf
• python
• qdrant
• rag
• react
• semantic-search
• typescript
• zero-trust-ai