sre-embedding-model 1.0.0

Fine-tuned SRE sentence embedding model — semantic search & similarity for Site Reliability Engineering

sre-embedding-model

Fine-tuned sentence embedding model for Site Reliability Engineering (SRE)

---

What is this?

A sentence embedding model fine-tuned on 636 SRE domain pairs across 16 domains. It understands SRE vocabulary out-of-the-box — no setup, no internet, weights bundled inside the package.

Base model: all-MiniLM-L6-v2 (Microsoft, Apache 2.0) — 22M parameters, 384-dim embeddings.

Trained on: Kubernetes pod states & errors · Observability & metrics · Incident management & RCA · APM & distributed tracing · Splunk & log analysis · Java / Spring Boot · MySQL · Kafka · API Gateway · CI/CD · CMDB · Cloud platforms · Security

---

Install

pip install sre-embedding-model

Model weights (~88 MB) are bundled — no internet connection required after install.

---

Quick Start

from sre_embedding_model import SREModel

model = SREModel()

# ── Similarity between two SRE texts ─────────────────────────────────────────
score = model.similarity("pod keeps crashing", "CrashLoopBackOff")
print(score)  # → 0.961

score = model.similarity("mean time to recover", "MTTR")
print(score)  # → 0.965

# ── Semantic search — rank docs by query ──────────────────────────────────────
runbooks = [
    "CrashLoopBackOff — container restart loop runbook",
    "OOMKilled — container memory limit exceeded",
    "Evicted — node resource pressure",
    "MySQL lock wait timeout — transaction blocked",
]
results = model.search("container is being killed due to memory", runbooks, top_k=3)
for doc, score in results:
    print(f"{score:.3f}  {doc}")
# → 0.924  OOMKilled — container memory limit exceeded
# → 0.871  CrashLoopBackOff — container restart loop runbook
# → 0.843  Evicted — node resource pressure

# ── Zero-shot classification ──────────────────────────────────────────────────
labels = ["Kubernetes error", "Database issue", "Kafka lag", "Deployment failure"]
results = model.classify("replica is not keeping up with the primary", labels)
print(results[0])  # → ("Database issue", 0.847)

# ── Encode to vectors (for vector DB / RAG) ───────────────────────────────────
vectors = model.encode(["MTTR", "SLO", "CrashLoopBackOff", "Kafka consumer lag"])
print(vectors.shape)  # → (4, 384)

# ── SRE relevance gate ────────────────────────────────────────────────────────
print(model.is_sre_related("pod is OOMKilled"))       # → True
print(model.is_sre_related("what is the weather?"))   # → False

# ── Rerank BM25/keyword results ───────────────────────────────────────────────
keyword_results = ["OOMKilled doc", "MTTR guide", "canary deployment playbook"]
reranked = model.rerank("container killed by kernel", keyword_results)

# ── Model info ────────────────────────────────────────────────────────────────
print(model.info())
# {
#   "embedding_dim": 384,
#   "max_seq_length": 256,
#   "base_model": "sentence-transformers/all-MiniLM-L6-v2",
#   "training_pairs": 636,
#   "domains": 16,
#   "license": "Apache-2.0"
# }

---

Singleton — for long-running services

from sre_embedding_model import get_model

# Loads once, reuses on every call — safe for SRE agent / FastAPI service
model = get_model()
score = model.similarity(alert_title, runbook_title)

---

API Reference

Method	Description
SREModel()	Load model (bundled weights)
model.encode(texts)	→ np.ndarray shape (N, 384)
model.similarity(a, b)	→ float cosine similarity
model.similarity_batch(anchors, candidates)	→ ndarray matrix
model.search(query, docs, top_k=5)	→ [(doc, score)] ranked
model.classify(text, labels)	→ [(label, score)] ranked
model.is_sre_related(text)	→ bool
model.rerank(query, docs)	→ all docs re-ranked
model.info()	→ metadata dict
get_model()	→ singleton SREModel

---

SRE Domains Covered

Domain	Examples
Kubernetes Pod States	CrashLoopBackOff, OOMKilled, ImagePullBackOff, Evicted, Pending
Kubernetes Resources	HPA, VPA, PVC, ConfigMap, Ingress, NetworkPolicy
Kubernetes Troubleshooting	kubectl describe, node pressure, resource limits
Observability	p99 latency, error rate, Prometheus, Grafana, alerting
Incident Management	MTTR, MTTD, RCA, postmortem, on-call, escalation
APM & Tracing	spans, traces, distributed tracing, Jaeger, OpenTelemetry
Splunk	SPL queries, HEC, indexes, sourcetypes, dashboards
Java / Spring Boot	NullPointerException, BeanCreationException, heap, GC
MySQL	deadlock, lock wait timeout, replication lag, slow query
Kafka	consumer lag, producer throughput, partition rebalancing
API Gateway	rate limiting, circuit breaker, upstream timeout
CI/CD	deployment strategies, canary, blue-green, rollback
Cloud Platform	EC2, EKS, ECS, auto-scaling, multi-AZ, RTO/RPO
Data Pipelines	Kafka streaming, ETL, dead letter queue
Security / SRE	mTLS, RBAC, secrets, zero-trust
Error Budget	SLO, SLA, burn rate, chaos engineering

---

Use Cases

• Alert deduplication — find semantically duplicate alerts
• Runbook retrieval — match incident description → correct runbook
• NL→SRE term mapping — "container keeps dying" → CrashLoopBackOff
• RAG / vector DB — encode SRE docs as dense vectors for retrieval-augmented generation
• SRE agent — semantic routing in LLM-powered SRE assistants
• Incident classification — classify incoming tickets by SRE domain

---

Performance

Metric	Base model	Fine-tuned
SRE similar pairs (in-distribution)	3/18 pass	18/18 pass
SRE similar pairs (out-of-distribution)	6/21 pass	21/21 pass
Mean cosine on SRE similar pairs	0.312	0.903
Val Spearman ρ	-0.263	+0.542

---

Requirements

sentence-transformers >= 2.2.0
numpy >= 1.21.0

---

License

Apache License 2.0 — see LICENSE

Base model all-MiniLM-L6-v2 © Microsoft Corporation, Apache 2.0. Fine-tuning, training data, and package © 2026 Senthil Kumar Thanapal.

---

Author

Senthil Kumar Thanapal · [e-mail]