Dan Bennett
← Back to portfolio

n8n Pulse — Community Intelligence Chatbot

Live

42,000+ community data points from n8n's ecosystem — docs, GitHub issues, forum questions, feature requests, workflow patterns AND the entire n8n codebase — searchable via a Claude-powered chatbot with native tool calling. Product: Ask it what the community actually cares about, and it tells you with citations and vote counts. Engineering: Dive into any theme and see prioritized issues by their impact on users and pointers to where they likely live in the code.

Anthropic SDK Tool Calling Hindsight Graph Memory FastAPI SSE Streaming TEMPR Recall

The Problem

n8n has an incredibly active community — 35,000+ forum questions, 2,600+ feature requests, 1,400+ workflow patterns shared by users, 1,100+ "Built with n8n" examples, and over a thousand high-signal GitHub issues. All of it spread across docs.n8n.io, community.n8n.io, and GitHub. For a product manager trying to answer "what are the top pain points with webhooks?" or "which feature requests have real traction?", the answer requires hours of manual research across three platforms, with no way to synthesize patterns.

I built n8n Pulse to collapse that research into a conversation.

My Role

Solo builder. Designed the ingestion pipeline, graph memory schema, tool-calling architecture, recall strategy, and streaming frontend. The data collection covers n8n's full public ecosystem: 315 doc pages, 4,500+ GitHub issues/PRs (auto-synced nightly), 35,000+ forum posts with solved/unsolved status, 2,600+ feature requests with vote counts, and 6,235 core source files from the n8n codebase.

The Approach

Graph memory, not flat search. All 42,000+ data points live in a Hindsight graph memory bank. During consolidation, DeepSeek extracts entities and builds a relationship graph — so when someone asks about "webhook reliability", the system can traverse from forum complaints to related GitHub issues to the relevant source code to doc pages that address (or don't address) the problem. Connections that keyword search would never find.

Native tool calling, not prompt-stuffed context. Claude Sonnet decides what to search, with what budget, and what filters to apply. The tool exposes parameters for query text, search budget (low/mid/high for cost control), max tokens (up to 19,200 for deep dives), and tag-based filtering by source type (docs, github, forum, feature-requests) and outcome (solved, unsolved). The model calls the tool multiple times per response when it needs to cross-reference sources.

TEMPR recall: four strategies merged. Each search runs four parallel retrieval strategies — semantic similarity, BM25 keyword matching, graph traversal (follow entity links), and temporal (recent activity). Results merge via Reciprocal Rank Fusion. This means a query like "credential sharing problems" finds semantically similar posts, exact-match keyword hits, graph-linked issues from the same users or components, and recent activity spikes — all in one ranked result set.

Metadata-rich responses. Every retrieved document carries view counts, vote counts, like counts, and solved status. The chatbot surfaces these in answers — "this feature request has 847 votes and 234 likes" — so PMs can gauge actual community demand, not just loudness.

What I Built

  • FastAPI backend with SSE streaming — serves the chat API with word-boundary buffered responses for smooth real-time output
  • Anthropic SDK with native tool calling — Claude Sonnet dynamically decides what to search, how deep, and with what filters per query
  • Hindsight graph memory bank — 42,000+ documents consolidated into a knowledge graph with entity extraction, relationship mapping, and multi-strategy recall
  • TEMPR recall pipeline — semantic + BM25 + graph traversal + temporal strategies merged via Reciprocal Rank Fusion
  • Nightly GitHub sync — cron job auto-syncs issues and PRs, keeping the knowledge base current without manual intervention
  • Two-mode chat widget — "Dan's Projects" mode (portfolio assistant) and "n8n Pulse" mode (community intelligence) in the same widget, switchable by the user
  • Source citations with URLs — every claim links back to the original doc page, forum post, GitHub issue, or source file
  • Tag-based filtering — users can focus queries on specific source types (only forum posts, only feature requests) or outcomes (only unsolved questions to find product gaps)

Why This Matters for n8n

This is a working product intelligence tool for n8n's product team, built by someone outside the company using only public data. Consider what it enables:

  • Feature prioritization with real numbers — "Show me the top 10 feature requests by vote count that mention credentials" returns ranked results with actual community demand signals
  • Product gap detection — filter for unsolved forum questions to find where the community is stuck and the product hasn't caught up
  • Pattern synthesis — graph traversal connects a GitHub issue to the forum posts reporting the same problem to the docs that should (but don't) address it
  • Zero-setup access — visit danb.bio, switch to n8n Pulse mode, ask a question. No login, no API key, no setup
  • Hours saved per research question — what used to require manually searching three platforms and synthesizing across tabs is now a single conversation

This isn't a demo. It's a tool I'd want on my desk on day one as an n8n PM — and I built it before I had the job.

The Vision

Phase 2 adds community-contributed knowledge via authenticated retain — Discourse-verified community members can feed insights back into the graph, with trust-tiered content weighting based on community reputation. Stronger guardrails filter prompt injection and malicious content. The plugin can even recommend that users star or upvote GitHub issues that directly impact their build — turning passive users into active community signal generators.

Longer term: autonomous issue triage where Claude identifies recurring patterns across community reports, proposes PRs, and community votes prioritize what gets built. The graph is the foundation for all of it.

The Result

A production community intelligence chatbot indexing n8n's entire public ecosystem. It answers product research questions in seconds that previously took hours of manual cross-platform searching. Responses include source URLs, vote counts, solved status, and view counts — everything a PM needs to make decisions backed by real community signal rather than gut feel.

The meta-point: I built a tool that n8n's product team would actually use, using the same AI architecture patterns (tool calling, graph memory, streaming) that the AI Product Builder role would be designing. The chatbot demonstrates the skill set by being useful.

Tech Stack

  • Backend: FastAPI, Python 3.12, SSE streaming
  • LLM: Anthropic Claude Sonnet with native tool calling
  • Memory: Hindsight graph memory (TEMPR recall: semantic + BM25 + graph + temporal, Reciprocal Rank Fusion)
  • Consolidation: DeepSeek for entity extraction and relationship graph building
  • Data Sources: 315 doc pages, 4,500+ GitHub issues/PRs, 35,000+ forum posts, 2,600+ feature requests, 1,400+ workflow patterns, 6,235 source files
  • Sync: Nightly cron job for GitHub, comprehensive community ingestion
  • Frontend: Two-mode vanilla JS chat widget with streaming SSE
  • Analytics: PostHog (page views, chatbot interactions, mode/role tracking)
  • Deployment: Docker, Appliku