Key facts
| Pattern | Corpus indexing → evidence retrieval → cited synthesis |
| Embeddings | plugsky-embed and plugsky-embed-multilingual via /v1/embeddings (live) |
| Synthesis | Long-context models for multi-document comparison (live) |
| Metadata | Publication date, author, study type and topic per chunk |
| Citations | Structured output with source IDs and quoted spans |
| Models | 30+ models behind one endpoint, route per stage |
| Deployment | Region-locked planes, VPC, on-prem and air-gapped options |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
TL;DR
- Curate the corpus — approved sources beat open-web crawling for defensible research.
- Store publication date and study type so recency and evidence quality can be judged.
- Retrieve evidence per sub-question, then synthesise across documents.
- Quote spans, not just links, so reviewers can verify claims quickly.
- Evaluate coverage and support, and track which sources never get retrieved.
How it works, step by step
- Define the corpus and inclusion rules: which publishers, study types and date ranges count as evidence.
- Parse documents with structure — abstract, sections, tables — and chunk by section with publication metadata.
- Embed chunks with plugsky-embed, or plugsky-embed-multilingual for multi-language research collections.
- Retrieve evidence per sub-question with filters for date, study type and topic.
- Add a verification pass that maps each drafted claim to a quoted span from the retrieved evidence.
- Synthesise with a long-context model, producing a brief with inline citations and a source list.
- Evaluate coverage, support and calibration on a fixed question set, and review which sources are never retrieved.
Try it yourself
Open the RAG architecture builder →
A research corpus instead of the open web
Open-web research agents are hard to defend: sources change, quality varies and citations rot. A curated corpus fixes that. You decide what enters the index, how it is chunked and what metadata travels with it, so every answer can be traced to a document you chose to trust.
The model layer is live on Plugsky: /v1/embeddings for indexing and /v1/chat/completions for synthesis, with long-context models for comparing many documents at once. The corpus, filters and permissions remain in your infrastructure.
Evidence tables and citations
Good research output is a table as much as a narrative: claim, supporting evidence, source, date and confidence. Building that structure during the workflow — rather than extracting it afterwards — makes verification routine and reveals gaps while there is still time to retrieve more.
- Quote spans: store the exact sentence supporting each claim for fast reviewer checks.
- Source quality: carry study type and publisher so synthesis can weigh evidence.
- Conflicts: surface disagreement between sources instead of averaging it away.
- Recency: flag claims resting only on older material.
Coverage, recency and evaluation
Evaluate a research pipeline on coverage and support, not eloquence. Coverage asks whether each sub-question found evidence; support asks whether every claim is traceable to a retrieved span; calibration asks whether the brief distinguishes strong from weak evidence. A fixed question set with analyst review keeps these honest.
Watch retrieval health too: sources that never appear in results may be poorly chunked, missing metadata or simply outside the questions asked. Refresh the index as new material arrives, and keep publication dates current so recency ranking stays meaningful. Route filtering and extraction to small models and reserve long-context synthesis for the final stage. Deploy in a region that matches your research data policy, including on-prem where corpora are restricted.
Honest comparison
| Concern | Curated-corpus RAG | Open-web agent | Manual literature review |
|---|---|---|---|
| Source control | You define inclusion rules | Whatever is crawled | Analyst selection |
| Citations | Quoted spans with IDs | Links that may rot | Notes and PDFs |
| Coverage checks | Sub-question retrieval scoring | Unmeasured | Manual |
| Freshness | Controlled re-indexing | Unpredictable | Depends on analyst |
| Audit | Traces of evidence per claim | Limited | Working files |
Frequently asked questions
Can the agent search the open web?
It can if you expose a web tool, but curated corpora produce more defensible research. Start with sources you control, then add approved external feeds deliberately.
How do we keep citations accurate?
Map every claim to a quoted span from retrieved evidence during the workflow, and store the source ID with it. Citations assembled after writing are usually wrong.
What metadata matters most?
Publication date, publisher or study type, and topic tags. These drive filters for recency and evidence quality, which reviewers care about more than similarity scores.
How large a corpus is practical?
Start with the documents needed for one research domain. Coverage of that domain matters more than volume, and incremental indexing lets the corpus grow as questions demand.
Which models handle synthesis?
Long-context models such as plugsky-longctx for comparing many documents, with smaller models for query generation and extraction. Everything sits behind one endpoint.
How should we evaluate the output?
Score sub-question coverage, claim support and calibration between strong and weak evidence, using a fixed question set reviewed by domain experts.
Can research corpora stay on-prem?
Yes. Plugsky supports region-locked cloud planes plus VPC, on-prem and air-gapped deployment, so restricted corpora can remain inside your environment.