Noetis Research · Business engine
Research Synthesis API for Supplied Documents
Produces a sourced brief answering a defined research question.
Answers: "Can an engine synthesise a source pack we provide and show where sources disagree?" For analysts and research teams.
Web search and retrieval are not available. The engine works only on the documents you send.
Available today
ImplementedImplemented: runs on the shared runtime with supplied material. One bounded development check on 2026-10-12 took 116.9s with 15 model calls and ended "incomplete". That is a development signal, not an evaluation. Customer access is by approved account.
What you send
- Question
- Supplied text material
- Text material up to 150,000 characters per run
What comes back
- Research brief with citations, disagreements and unresolved questions.
- Findings with the located quote and a support status from claim review
- Release status: released, incomplete with reasons, or failed
- Executed roles, usage and measured cost
Sample
Authored illustration. Written by Encevra to show the workflow. Not customer data and not a recorded model output.
What do the supplied studies say about four day work weeks and productivity?
Source A (2023 UK pilot, 61 firms, self-reported): 92% continued; revenue roughly flat. Source B (2019 single firm press release): productivity up 40%, no method stated. Source C (2024 survey, 1,000 workers): 63% prefer it; productivity not measured.
- Points of agreement across the supplied sources, each with a quoted passage
- Disagreements stated side by side with both passages
- Unresolved questions the sources do not answer
The real response shape is documented in the API reference.
How the workflow operates
Each role is a responsibility in the workflow, served by a third party language model. More roles are not automatically better; depth and routing decide which run.
Built with
Models this engine uses: OpenAI (GPT), Anthropic (Claude). Specialists run on OpenAI models and the claim review and final brief on Anthropic models. If a provider has a temporary failure, the step can fall back to an OpenAI model, and the run records that.
Research Planner
Define scope, subquestions and stopping conditions.
Source Researcher
Retrieve relevant sources using available, authorized tools.
Source Quality Reviewer
Check provenance, freshness and relevance.
Contradiction Analyst
Investigate conflicting findings and alternative explanations.
Coverage Reviewer
Identify gaps and check conclusions answer the question.
Evidence and deployment
Evidence
This engine reads the text you send in the request. Each released finding carries the exact quote and its character span in that text, located mechanically before release. Quotes that cannot be located are labelled Insufficient evidence and never shown as source verified. Interpretation is reviewed as a separate step, and gaps such as a missing clause describe what was searched instead of citing text that is not there.
Deployment
Runs on Encevra hosted infrastructure using the third party models listed above. Every model call and fallback is checked against the deployment policy first. Private cloud and air-gapped deployment are planned; discuss deployment requirements.
What it does not handle
- Questions requiring sources the engine cannot access
- Authorized source retrieval is not available: Not built. Engines read only supplied material.
- Uses only supplied material. No freshness guarantee beyond the documents provided.
- Required for every output during trials.
How to integrate
Send "engine": "research" with your question and material from your server. Keys stay server side.
How to evaluate it
No published result for this workflow yet. Build a small answer key from your own work and compare against a strong single model baseline. See the evaluation checklist and method.
Cost and rights: pricing and licensing. Showing outputs to your own customers needs a written agreement.
Related engines
Content reviewed 2026-10-12 by the Encevra engine team. Engine version 0.1.0 on noetis-core-1.