sentence-transformers How to use Akshayram1/embeddinggemma-300m-v2 with sentence-transformers:
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("Akshayram1/embeddinggemma-300m-v2")
sentences = [
"what is still blocking dead letter replay tooling",
"Dead Letter Queue replay tooling\n\nOperators need a safe way to replay events out of the Dead Letter Queue after a fix ships. Add atlasctl dlq replay backed by api/pipelines.py with filters on pipeline, failure class and time window, a dry-run mode that prints counts only, and a rate limit so replay never starves live traffic. Acceptance criteria: replayed events keep their original event_id so downstream deduplication still holds, every replay writes an audit record, and the runbook contains a worked example from the 2026-08-26 regression. Status history: todo 2026-08-25, in_progress 2026-09-02, in_review 2026-09-09 with two open review comments from Vikram Joshi about rate limiting. Depends on ATLAS-107, ATLAS-108 and ATLAS-116. Reporter Vikram Joshi, assignee Neha Kapoor.",
"INC-002 Staging Ingestion Outage from SSL Verification Failure\n\nDocument kind: incident. Author: Vikram Joshi. Reviewers: Rohan Mehta, Maya Rao, Akshay Sharma. Status: active, written 2026-08-19, reviewed at the postmortem on 2026-08-21. Impact: all staging CDC ingestion down for 4 hours 11 minutes, 2026-08-19T06:12Z to 10:23Z. No customer data was affected; the pilot demo scheduled for that afternoon was postponed. Timeline: 06:12Z connector pods begin crash-looping with certificate verify failed; 06:40Z Vikram Joshi opens ATLAS-110 and states on the bridge that the staging PostgreSQL server certificate has expired; 07:20Z a certificate renewal request is filed with the platform team, which is the wrong remediation; 09:05Z Rohan Mehta checks openssl output and finds the server certificate valid until 2027-01-14; 09:30Z he execs into a pod and finds /etc/ssl/certs/aster-internal-ca.pem absent; 09:40Z ATLAS-111 is opened, tracing the loss to the slim base image adopted on 2026-08-14; 10:23Z a rebuilt image restores ingestion. CONTRADICTION RESOLVED: the expired-server-certificate claim recorded in the first version of ATLAS-110 is incorrect and superseded by ATLAS-111. Corrective actions: CA bundle baked in with a build-time fingerprint assertion and a start-up self-check, per the SSL hardening decision, shipped in PR-004 with commit e42a7f0c3b91; the SSL certificate runbook now mandates checking the client side before the server. 90 minutes of the outage are directly attributable to diagnosing in the wrong order.",
"ADR-004 Retry Budget and Backoff Policy\n\nDocument kind: decision. Author: Vikram Joshi. Approvers: Neha Kapoor, Rohan Mehta, Maya Rao. Status: approved 2026-08-26, implemented 2026-09-01 under ATLAS-108. Context: the first draft of the DLQ replay runbook stated a retry budget of five attempts with a 30 second ceiling. During the retry and DLQ design review Vikram Joshi pointed out that Aster Labs broker maintenance routinely takes eight to twelve minutes, so a 30 second ceiling would exhaust every retry and dump an entire maintenance window into the Dead Letter Queue, turning a self-healing event into a manual replay of tens of thousands of records. Neha Kapoor countered that a long ceiling delays the visibility of genuinely broken pipelines. Decision: eight attempts with exponential backoff and full jitter, base 2 seconds, ceiling 15 minutes, giving roughly 45 minutes of tolerance; and, to answer the visibility concern, a pipeline with any event past attempt four for more than five minutes raises a warning on the Observability Dashboard under ATLAS-119. The earlier five-attempt, 30-second text in the runbook is superseded and the runbook was corrected on 2026-08-27. Implemented in workers/retry_worker.py through PR-005 and commit b8740e2fd1a3, with ceilings configurable in deploy/helm/atlas/values.yaml."
]
embeddings = model.encode(sentences)
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [4, 4] 
