Siggy learns from every incident and fix. Three-layer retrieval makes recommendations sharper over time. Built on SigNoz.
Modern observability gives you data. Siggy gives you answers.
Teams drown in hundreds of alerts per day. Most are noise. The critical ones get buried because there is no historical context to prioritize them.
The fix for that Redis timeout lives in one engineer's head. When they are offline, the team restarts services blindly and hopes. Knowledge walks out the door every Friday.
Mean time to resolution has not improved in years because every investigation starts from scratch. No system remembers what worked before.
Three-layer retrieval, explainable recommendations, and a feedback loop that gets smarter with every incident.
Vector search via Qdrant, knowledge graph via SQLite, and Bayesian experience ranking. Three retrieval layers combined into one pipeline that improves over time.
5 node types, 6 relationship types. Maps services, failure modes, and fixes into a connected graph that expands context during investigation.
Records outcomes, computes success rates, and uses Bayesian ranking so the best fixes rise to the top. Every resolved incident makes future recommendations better.
siggy instrument python app.py wraps any Python app with OpenTelemetry. Zero code changes. One command.
Every recommendation includes root cause, fix, confidence score, evidence chain, and reasoning. Not a black box -- a transparent assistant.
pip install siggy-memory then siggy quickstart. Embedded Qdrant (no Docker), local embedding fallback, no LLM required for normalization.
Auto-creates 10-panel dashboards, 3 default alert rules, saved views, and webhook notification channels. Alert sidecar enriches in real time.
Polls SigNoz alerts every 30 seconds, enriches with memory pipeline, and writes recommendations back as OTel spans visible in your SigNoz dashboard.
siggy demo seeds 4 realistic incidents with outcomes. Works without SigNoz, without Docker, without anything but Python.
Seven steps from alert to explainable recommendation.
SigNoz detects an anomaly in your traces
Incident becomes a structured KnowledgeObject
Qdrant finds similar past incidents
Knowledge graph adds related services and patterns
Bayesian scoring by historical outcomes
Root cause, confidence, evidence chain
Actionable fix written back to SigNoz
Engineers record outcomes. The experience engine learns. Future recommendations improve.
Every query flows through three memory layers. Each layer contributes signals. The result is ranked by real-world outcomes.
Semantic similarity across all past incidents. Finds the closest matches by embedding distance.
Expands context through service dependencies, failure types, and fix relationships. 5 node types, 6 relationship types.
Ranks candidates by historical success rate, resolution time, and engineer feedback. Best fixes rise to the top.
Siggy does not just connect to SigNoz. It sets it up for you.
10 panels covering error rates, latency percentiles, memory hits, graph depth, experience scores, and recommendation throughput.
3 alert rules configured out of the box. Catches high error rates, latency spikes, and high failure counts across all services.
Polls SigNoz alerts every 30 seconds, enriches through the memory pipeline, and writes recommendations back as OTel spans.
Pre-configured saved views for incident investigation. Webhook notification channel for sending recommendations to Slack, PagerDuty, or any endpoint.
A CLI that works like a senior SRE sitting next to you.
Battle-tested taxonomy, real benchmarks, measurable performance.
Three paths. Pick the one that fits.
Embedded Qdrant, auto-detects SigNoz and GROQ. Configures everything including dashboard and alert rules.
Full control over each connection. Configure SigNoz URL, Qdrant endpoint, and LLM provider individually.
Seeds 4 realistic incidents. No SigNoz, no Docker. Just Python 3.10+.
Requires Python 3.10+ and pip