Open Source: Dealflow Data Platform – AI + Neo4j Integration

[santosh33]

Hi everyone

I’m Santosh Narayanan, a Senior Full Stack & Cloud Engineer specializing in Generative AI and LLM Systems.

I recently published the Dealflow Data Platform, an open-source project that integrates LangChain, Neo4j Aura (free tier) / Neo4j Desktop, and Google Cloud Run to explore AI-driven graph reasoning and Cypher query generation.

This post marks the completion of Phase 1, where I focused on building the AI–graph integration layer and end-to-end workflow for intelligent data mapping.
Phase 2 will expand into deeper reasoning, richer data models, and interactive visualizations.

GitHub: https://github.com/santoshnarayanan/dealflow-data

More info one can view from README.md and the Technical-design.md

[santosh33]

Example continuation post (Phase 2 → Phase 4 + Phase 5 preview)

Continuing my progress on the AI-powered Dealflow Dashboard (LangChain + Neo4j + Weaviate + React)!
Since the last update, I’ve completed Phase 2, Phase 3, and Phase 4, expanding the system into a full hybrid RAG + multi-agent reasoning platform. Here’s the gist

---

Phase 2 — Vector DB + Semantic Search (Weaviate)

• Added Weaviate with text2vec-openai for embeddings.

• Created startup & investor vector schemas.

• Implemented semantic search endpoints for:

  • /vector/startups?q=...

  • /vector/investors?q=...

• Frontend semantic search UI to explore AI-powered similarity results.

This enabled contextual RAG workflows and improved natural-language discovery.

---

Phase 3 — Hybrid AI (Graph + Vector + LLM Synthesis)

• Implemented 3 major LangChain pipelines:

  • askGraph → LLM → Cypher → Neo4j

  • askRag → Semantic retrieval via Weaviate

  • askHybrid → Combine Neo4j + Weaviate + LLM summarization

• Frontend pages added:

  • AI Query, Semantic Search, Hybrid AI

• Architecture now supports richer multi-source reasoning.

The system can now answer queries using BOTH graph relations and semantic context.

---

Phase 4 — Multi-Agent Orchestrator (Classifier → Cypher → Vector → Answer Agent)

Built a production-style multi-agent reasoning workflow:

Classifier Agent — routes user queries to cypher, vector, or hybrid
Cypher Agent — generates Cypher + executes on Neo4j
Vector Agent — semantic search on Weaviate
Answer Agent — synthesizes JSON answer + explanation

Additional upgrades:

• Prometheus metrics (/metrics)

• Detailed pino logging

• Health checks for Neo4j, Weaviate, Agents (/health/*)

• New frontend module: AI Multi-Agent Chat

The system now intelligently selects reasoning paths and explains how it arrived at answers.

---

What’s coming in Phase 5 (preview)

Phase 5 is focused on production-grade deployment + observability + system intelligence:

Deployment (GCP)

• Cloud Run services (frontend + backend)

• Managed vector DB option (Weaviate Cloud / Vertex AI Vector Search)

• Traefik / API gateway routing

Observability

• OpenTelemetry distributed tracing

• Grafana dashboards for metrics

• Slow-query monitoring for Neo4j & Weaviate

• Error budget & performance thresholds

Multi-Agent Enhancements

• Short-term + long-term memory (Redis)

• Agent-level caching to reduce LLM calls

• More domain agents (Funding Rounds, Investor Profiling)

Graph Expansion

• Adding FundingRound relationships

• Deeper investment patterns & reasoning queries

---

Repo: https://github.com/santoshnarayanan/dealflow-data