A single chat completion can draft text. Production workflows need routing, tools, memory, and clear ownership of failure. Multi-agent systems earn their complexity when the work naturally splits: search vs finance vs market research; plan vs code vs execute.
\nThis post captures patterns from shipping:
\nLangGraph-style orchestration gives you:
\nThat beats an unbounded “agent loop” that mutates a blob of chat history until tokens run out.
\nA durable shape:
\nUser → Supervisor / orchestrator\n ├─ Specialist A (tools)\n ├─ Specialist B (tools)\n └─ Synthesizer → final artifactKeep the supervisor thin: classify intent, pick agents, merge results, enforce budgets (steps, tokens, wall clock).
\nIn the real-estate platform, separate agents for search, finance, and market work better than one prompt that tries to do everything:
\nThe Agentic Data Analyst follows the same idea along a pipeline: understand query → generate code → sandbox execute → present in Streamlit. Code execution is a different trust boundary than “write an answer”; isolate it.
\nCommon failure modes (and mitigations):
\n| Failure | \nSymptom | \nMitigation | \n
|---|---|---|
| Tool hallucination | \nInvented APIs / paths | \nStrict tool schemas; reject unknown names | \n
| Infinite loops | \nSame node retries forever | \nMax steps; circuit breakers; escalate to human | \n
| Context bloat | \nLatency and cost explode | \nSummarize intermediate state; pass IDs not dumps | \n
| Unsafe code | \nArbitrary exec in analyst flows | \nDocker/host sandbox (e.g. patterns like codibox); no network by default | \n
| Silent wrong answers | \nConfident nonsense | \nCritic / grader node; require citations or tests | \n
Design the graph so failure is a first-class edge, not an exception you only see in logs.
\nFinancial and analytical users ask “how did you get this?” Checkpointed graphs let you:
\nMulti-LLM routing belongs in the orchestrator, not scattered inside each agent.
\nSkip the swarm if:
\nStart with a single agent + tools. Promote to multi-agent when handoffs and specialization show up in the eval set, not in a slide deck.
\nMulti-agent orchestration is software architecture with LLMs in the loop — treat it like distributed systems, not magic.
"}},{"node":{"frontmatter":{"title":"GraphRAG in Production — Hybrid Retrieval That Auditors Can Trust","description":"How Neo4j knowledge graphs plus vector and full-text retrieval produce explainable answers for financial due diligence — patterns from building a GraphRAG knowledge engine.","slug":"/pensieve/graphrag-production","date":"2026-03-01","tags":["GraphRAG","Neo4j","RAG","LLM","Production"],"draft":false},"html":"Vector-only RAG is great at “find something that sounds like this.” Due diligence and compliance need something sharper: who owns what, which clause constrains which entity, and why an answer is grounded in a specific filing.
\nEmbeddings alone blur those edges. Entities collapse into similar neighborhoods; cross-document relationships disappear; citations become soft suggestions instead of audit trails. That is the gap GraphRAG is meant to close.
\nA practical GraphRAG stack for financial and compliance corpora looks like this:
\nThe graph is not decoration. It is the index that makes “explainable” possible: you can show the path from claim → entity → document span.
\nIn practice, three channels complement each other:
\n| Channel | \nStrength | \nFailure mode | \n
|---|---|---|
| Vector | \nParaphrase and conceptual match | \nMisses exact identifiers, clauses, ticker symbols | \n
| Full-text / lexical | \nExact terms, IDs, regulation codes | \nBrittle to wording changes | \n
| Graph expansion | \nMulti-hop ownership, parties, obligations | \nNeeds decent extraction quality | \n
A solid fusion strategy ranks candidates from vector + FT, then expands 1–2 hops in Neo4j (related entities, linked clauses, prior filings). Rerank before the LLM so the context window stays dense and cheap.
\nProduction buyers in finance care less about flashy demos and more about:
\nTreat the LLM as a reasoner over evidence, not as the system of record.
\nThese patterns underpin the GraphRAG Knowledge Engine featured on this site: Neo4j knowledge graphs, hybrid vector and full-text retrieval, and LLM reasoning aimed at due diligence and risk analysis. Client-facing deployment details stay private until the customer environment is public; the architecture above is the portable part.
\nIf you are evaluating GraphRAG vs “bigger context window,” start with retrieval quality and provenance. Model upgrades are easy to swap. A trustworthy evidence layer is not.
"}},{"node":{"frontmatter":{"title":"Sadeeq Al-Siha Medical Chatbot","description":"Medical RAG QA with Meditron 7B, Qdrant, and PubMedBERT for evidence-grounded clinical answers","slug":"/projects/sadeeq-al-siha","date":"2024-01-15","tags":["LLM","Chatbot","Healthcare","RAG"],"draft":false},"html":"Medical RAG (Retrieval-Augmented Generation) QA application using the Meditron 7B LLM, Qdrant vector database, and PubMedBERT embeddings — designed for accurate, literature-grounded medical answers through a Chainlit interface.
\nRepo: oelbourki/Sadeeq-Al-Siha-Medical-Chatbot
"}},{"node":{"frontmatter":{"title":"Ads-LLM","description":"A program for matching articles with ads using advanced NLP techniques","slug":"/projects/ads-llm","date":"2023-12-01","tags":["LLM","Advertising","NLP"],"draft":false},"html":"Implemented a program for matching articles with ads, utilizing KeyBERT to pinpoint relevant article keywords and a custom LLAVA deployed on Vertex AI to extract ad-descriptive keywords.
\n