citation-pulsar-helm 🌟

Helm chart for deploying a local LLM on Kubernetes via llama.cpp

Pulsar is the inference engine behind the citation-cosmograph project. It deploys a quantized language model as an OpenAI-compatible API endpoint on Kubernetes, providing the steady signal that powers Astrolabe's venue governance extraction pipeline.

pulsar 🌟  →  astrolabe 🔭  →  citation-constellation ✨
(the signal)    (the instrument)    (the map)
    ▲
  you are here

Academic venue governance data — who sits on which editorial board, who serves on which program committee — is scattered across thousands of heterogeneous journal and conference websites. Extracting this data at scale requires a language model capable of structured information extraction from raw HTML. Pulsar provides that model as a self-hosted, private, Kubernetes-native service — no commercial API keys, no data leaving the cluster, no per-token costs.

Pulsar will deploy a llama.cpp server running a quantized open-weight language model, exposed as an OpenAI-compatible /v1/chat/completions endpoint within the Kubernetes cluster.

Target model: Qwen 3.5 8B (Q4_K_M quantization) — selected for its strong structured extraction capabilities at a size that runs on CPU-only nodes.

Key design goals:

• CPU-only inference. No GPU required. The quantized model will run on standard Kubernetes worker nodes with 4–8 vCPU and 24 GB RAM, making it deployable on typical academic computing infrastructure.
• OpenAI-compatible API. Astrolabe and any other consumer will interact with Pulsar through the standard /v1/chat/completions endpoint, making it straightforward to swap models or inference backends in the future.
• Cluster-internal only. The API endpoint will not be exposed externally. Pulsar is a backend service consumed by Astrolabe within the same cluster.
• Persistent model storage. Model weights will be stored on a Kubernetes PersistentVolume, avoiding re-download on pod restarts.
• Health monitoring. Liveness and readiness probes will ensure the model is loaded and responsive before Astrolabe sends extraction requests.

# Add the Helm repo (once published)
helm repo add citation-cosmograph https://citation-cosmograph.github.io/citation-pulsar-helm

# Install with default values
helm install pulsar citation-cosmograph/citation-pulsar-helm

# Or install from local chart
helm install pulsar ./charts/citation-pulsar-helm

The Helm chart will support configuration of model selection, quantization level, resource limits, replica count, and storage class via values.yaml.

Pulsar exists to serve Astrolabe. The typical flow will be:

1. Astrolabe identifies a venue whose editorial board page needs extraction.
2. Astrolabe fetches the raw HTML (via httpx or Cloudflare Crawl API).
3. Astrolabe sends the HTML to Pulsar's /v1/chat/completions endpoint with a structured extraction prompt.
4. Pulsar returns a JSON array of governance members (name, role, institution, ORCID).
5. Astrolabe performs entity resolution and stores the results.

Pulsar does not know or care about citation analysis. It is a general-purpose LLM inference service that happens to be optimized for the structured extraction workload Astrolabe requires.

🔧 In development. The Helm chart, default model configuration, and smoke tests are being prepared. This repository will contain the chart, documentation, and validation scripts once the initial version is ready.

For the full methodology, conceptual foundations, tool landscape comparison, discussion of responsible research assessment alignment, and detailed limitations analysis, see the accompanying research paper:

Mahbub Ul Alam. Where do your citations come from? Citation-Constellation: A free, open-source, no-code, and auditable tool for citation network decomposition with complementary BARON and HEROCON scores, 2026. URL: https://arxiv.org/abs/2603.24216, arXiv:2603.24216, doi:10.48550/arXiv.2603.24216.

The paper is also available embedded within the web tool under the Full Research Paper tab.

@misc{alam2026citationconstellation,
  title={Where Do Your Citations Come From? {C}itation-{C}onstellation: A Free, Open-Source, No-Code, and Auditable Tool for Citation Network Decomposition with Complementary {BARON} and {HEROCON} Scores},
  author={Mahbub Ul Alam},
  year={2026},
  eprint={2603.24216},
  archivePrefix={arXiv},
  primaryClass={cs.DL},
  url={https://arxiv.org/abs/2603.24216},
  doi={10.48550/arXiv.2603.24216}
}

Powered by OpenAlex, ORCID, and ROR.

MIT