Frequently
Asked

Polari Technologies builds data pre-processing infrastructure for applications that use large language models. The core idea is simple: most of what gets sent to expensive frontier models — extraction, classification, deduplication, clustering — is work that smaller, purpose-built models do faster, cheaper, and more consistently. We do that work first, so the large model gets structured input instead of raw noise.

The Polari API is the developer-facing product. Digestr is our consumer news application, built entirely on the same pipeline. Both run on the same infrastructure.

Polari Technologies LLC is the company. The Polari API (polariapi.com) is our developer product — the processing pipeline available over HTTP. Digestr (digestr.ai) is our consumer news application. Eventide is our live graph visualization. All of them run on the same underlying infrastructure.

Anyone who has a corpus of documents, feeds, or records they want to make intelligently queryable — without spending months building the extraction and structuring pipeline themselves.

In practice that includes developers building LLM-powered applications, teams that need media monitoring or competitive intelligence, researchers working with large document sets, and organizations that have internal data sitting idle because the infrastructure to process it hasn't been worth building from scratch.

News is our primary and deepest implementation. But the pipeline — embed, extract, cluster, graph — applies wherever the relationships between documents matter as much as the documents themselves.

No. News is where the pipeline is most developed and battle-tested — we've been running it in production for two years — but the architecture doesn't care what the data is. The same four stages that process news articles can handle legal filings, research papers, support tickets, financial reports, internal communications, or any other text-heavy corpus where you want structure out of prose.

If you have a use case outside news, reach out. We're actively interested in expanding what the pipeline handles.

The pipeline runs in four stages. Each one builds on the output of the last:

• Layer 0 — Ingestion and quality. Raw content comes in. L0 scores it, filters low-signal material, and generates 384-dimensional semantic embeddings. Around 80% of raw input doesn't continue. That's by design.
• Layer 1 — Extraction. Everything that passed L0 gets analyzed at the sentence level. Named entities are extracted — people, organizations, locations, events. Sentiment is scored per article and per entity. Sentence-level embeddings are generated for fine-grained semantic search.
• Layer 2 — Clustering. Related documents are grouped into unified clusters using cosine similarity over embeddings combined with entity overlap. This is semantic equivalence detection, not keyword matching. Precision: 92%+. Deduplication: 60%+.
• Layer 3 — The graph. Entity co-occurrence across all clusters builds a weighted relationship graph. Narrative lifecycle states are assigned. The graph rebuilds every six hours. This is what Eventide renders, and what gives an LLM genuine relational context rather than a flat document list.

Starter tier includes Layers 0–2. Layer 3 requires Professional or Enterprise.

Register at polariapi.com. Free trial accounts receive a key immediately with 1,000 included API calls and access to all four layers for 14 days — no credit card required. Paid tier keys are issued automatically on subscription.

All requests authenticate via the Authorization: Bearer YOUR_API_KEY header.

English across all four layers. Spanish, French, German, and Chinese are on the 2026 roadmap. If a specific language is urgent for your use case, reach out — we can discuss enterprise arrangements.

Enforced per API key by tier:

• Starter: 1,000 requests/day, max 10/minute
• Professional: 10,000 requests/day, max 100/minute
• Enterprise: custom limits with burst support

All responses include X-RateLimit-Limit, X-RateLimit-Remaining, and X-RateLimit-Reset headers. Exceeding limits returns 429 Too Many Requests.

Most endpoints respond in under 200ms. You're querying pre-processed structure, not triggering live extraction — that distinction is why the latency is low. Clustering and graph queries may run up to 500ms depending on result size. Enterprise tiers have access to priority processing queues.

Yes, on Professional and Enterprise tiers. Subscribe to events including new cluster creation, trend detection, entity mention spikes, and sentiment shifts. Professional supports up to 5 active webhooks; Enterprise has no limit.

See the webhook documentation for payload schemas and setup instructions.

Annual billing saves 20%. Overages are billed at $0.001 per call above your plan limit. Full pricing at polariapi.com/pricing.

Yes. 14 days, 1,000 API calls, access to all four layers. No credit card required. At the end of the trial your key deactivates unless you subscribe.

Yes. Cancel from your account dashboard at any time. Access continues until the end of the current billing period. No prorated refunds for partial months on monthly plans; no contracts.

Annual plans are non-refundable except where required by applicable law.

Additional requests above your plan limit are billed at $0.001 per call. You can set a hard cap or spending alert in your account dashboard. Overages are invoiced at end of billing cycle.

Digestr is a news reader built entirely on the Polari pipeline. Articles come in as raw text; what surfaces to the reader is structured — clustered by story, deduplicated across sources, ranked by substance rather than recency.

It's also the reference implementation for the API. If something works well in Digestr, it's because the underlying pipeline is working. Two years in production, real users. That's the honest test we apply before opening anything to developers. The app lives at digestr.ai.

Digestr is in open beta. Core features are free. Premium features — custom topics, deeper history, advanced filters — will come with a paid tier at launch. If you're using Digestr now, you'll be grandfathered into a reasonable plan when billing goes live.

The Polari pipeline clusters articles from dozens of sources into unified stories. Within each cluster, Digestr surfaces the most informative and highest-quality coverage — not just the most recent. Reading patterns influence ranking over time.

No ads. No advertiser influence on ranking. We don't sell reading data.

Everything in Professional, plus:

• Custom API call volume with burst support
• 99.9% uptime SLA
• Dedicated account manager
• Phone and Slack support
• 90-day+ data retention
• Data Processing Agreement for compliance requirements
• Opt-out from model improvement programs
• White-label deployment options
• Custom source ingestion — bring your own data
• Custom integrations on request

Enterprise starts at $2,000/month. Contact sales@polaritechnologies.com.

Yes, on Enterprise agreements. If you have an existing corpus — internal documents, proprietary feeds, archived records — we can ingest and process it through the same pipeline. Your data is isolated in its own namespace and never mingles with other customers' data or the shared news corpus. Contact us to discuss scope and fit.

Yes, on Enterprise agreements. White-label options let you deploy the pipeline under your own brand. Pricing depends on scope — contact us to discuss.

Enterprise customers receive a 99.9% uptime SLA. Starter and Professional tiers operate on a best-effort basis without a formal guarantee, though we aim for equivalent reliability in practice. Historical uptime is published at our status page.

We use small, purpose-built models for each task — not a general-purpose frontier model doing everything. Embeddings are generated with BAAI/bge-small-en-v1.5 and all-MiniLM-L6-v2. Entity extraction uses spaCy. Sentiment scoring uses cardiffnlp/twitter-roberta-base-sentiment-latest. Clustering runs on our own two-stage algorithm built on top of these embeddings.

A large model (Claude Haiku, with Deepseek as fallback) is used only at Layer 3 for cluster title generation — the one task that genuinely benefits from language generation rather than classification or similarity. Everything else in the pipeline runs without a frontier model call.

This keeps the pipeline fast, deterministic, and not exposed to third-party rate limits or pricing volatility on the critical path.

You can. The results will be slower, more expensive, and less consistent — especially at scale. Embedding, entity extraction, clustering, and sentiment scoring are well-understood tasks that small, specialized models handle better than a general-purpose model does. They're faster by an order of magnitude, cost a fraction of the price per document, and produce more deterministic output.

The right pattern is: small models handle structure, large models handle reasoning. Polari is the structure step. The large model you're using gets clean, pre-extracted input and can focus on what it's actually good at.

100+ sources currently, via RSS feeds, direct API integrations, and web monitoring. Coverage skews toward English-language publications across news, technology, finance, and policy. Enterprise customers can request specific source additions or bring their own content for processing through a private namespace.

Python and JavaScript SDKs are on the Q1 2026 roadmap. The API is a clean REST interface that works with any HTTP client in the meantime. The documentation includes code samples in Python, JavaScript, and cURL for every endpoint.

For bugs or unexpected behavior, email hello@polaritechnologies.com or open a ticket through your account dashboard.

For security vulnerabilities, disclose responsibly to security@polaritechnologies.com before any public disclosure. We commit to acknowledging receipt within 48 hours and resolving confirmed issues within 30 days where possible.