Learn · AI methods
Local models, agents, and orchestration for research
The concept map showed what each AI idea does and does not establish. This page goes deeper on the three that matter most for a private, human-first workflow: running a model on your own machine, giving it tools, and coordinating those tools. At every step, the judgment stays with you.
After this guide you will be able to:
- Say what "local model", "agent", and "orchestration" mean, and why a researcher would want each.
- Split a workflow into steps a machine may run, steps it may only propose, and steps a human must decide.
- See why a private, agentic, well-orchestrated system is still not an "AI researcher".
On this page
Local models what · why · how
What
A local model is an AI model that runs on your own computer instead of a company's server, with no cloud call, no API key, and no account. Runners like Ollama and llama.cpp now put open models in the 7-to-8-billion-parameter class within reach of a modern laptop, and smaller task-specific models run on almost anything.
Why, for research
- Privacy by construction. Interview transcripts, unpublished manuscripts, and sensitive records never leave your device, because there is nowhere for them to go. This is the difference that makes AI ethical to use on interview data and lets you transcribe offline.
- No telemetry, no lock-in. Nothing is logged to a vendor; you are not one policy change away from losing access.
- Reproducibility. You can pin the exact model and version and record it, so an analysis can be re-run and described honestly in a methods section.
- Structure can substitute for scale. A July 2026 study of a control framework called Atomic Task Graph (arXiv:2607.01942) tested three interactive benchmarks. With an explicit task structure, 7-8B local models matched or beat the same models under weaker control. On some tasks they also exceeded a much larger model using a simpler method. The finding is narrow and useful: a local model becomes more capable when its work is bounded, decomposed, and checked, which is exactly how a research tool should use one.
How
Install a runner, pull a model, and point it at one bounded question at a time: draft first-cycle codes, suggest which passage of a source bears on a claim, flag inconsistent terminology. The model answers a narrow local question; the record and the decision stay with you.
What it does not do: decide truth. A local model is still a language model, and fluent output is not supported output. Our own laptop test found the honest limits: can a local model check your citations?
Agents what · why · how
What
An agent is a model given tools and a loop: it decides an action, takes it, observes the result, and decides the next one. With the right tools it can retrieve a source, run an analysis, and write out a table, pursuing a goal over several steps rather than answering a single prompt.
Why, for research
Real research work is multi-step, and the steps depend on each other. An agent can carry a chore across those steps (fetch, extract, compute, format) without you hand-copying between tools. This is useful for the mechanical parts of a workflow.
How: the part that protects the work
The safe way to use an agent in research is to decide, in advance, what it is allowed to do on its own. Sort every step into three kinds of node:
Deterministic, checkable work: parse a citation file, compute an effect estimate, render a plot, hash a protocol, compare two files.
Judgement-shaped work offered as a suggestion: classify a study design, flag a citation that may not support its claim, propose assumptions to check. Output is a proposal or a warning, not a verdict.
Consequential calls that block what comes after them: approve the research question, accept an inclusion decision, judge risk of bias, sign off on an interpretation. These wait for a qualified person.
What it does not have: authority. Being able to perform a step is not the same as being entitled to decide it, and only the second protects a research record. The agent proposes and does the mechanical work; a person disposes and signs.
Orchestration what · why · how
What
Orchestration is coordinating several bounded tools and agents over a graph (a defined order of steps with their dependencies) instead of asking one large model to do everything at once. Think of it as a workflow diagram the system actually runs, not a single "do my research" prompt.
Why, for research
- Small, replaceable, inspectable beats a monolith. When each step is a bounded tool, you can see what it did, swap a better one in, and improve one part without rebuilding the whole system.
- Independent steps can run in parallel, and when one step fails you can repair that part rather than restarting everything.
- Change becomes visible. An orchestrated research graph can answer: "If this changes, what must I look at again?" When an upstream object changes, such as a redefined construct or a retracted source, the system can trace its dependencies. It can then flag the specific downstream items that need review.
How
Under the hood this needs three things working together: a place to keep the graph (memory), a way to retrieve from your own material (RAG), and a way to retrieve along relationships (GraphRAG). The orchestration layer walks that structure and routes each result to the right node type: automatic for executable steps, a flag for assessments, and a gate for human decisions.
What it does not guarantee: that the workflow is scientifically appropriate. A perfectly coordinated graph can still be built around the wrong question. Orchestration keeps the work orderly and its consequences traceable; it does not make the design sound.
Where memory, RAG, and GraphRAG fit
These three are the machinery that makes local orchestration worth having, each closing a gap the one before it leaves:
- Memory is a persistent record of the work, the graph of objects and decisions, so the workflow, not a disappearing chat log, is the thing you keep and can audit.
- RAG (retrieval-augmented generation) lets a model answer from your material (your vault, your papers) instead of its training data alone, and cite what it used.
- GraphRAG retrieves along the relationships between objects (this claim, the source it cites, the analysis that source rests on) so questions about how things connect can be answered, which plain chunk retrieval does not reliably do.
Put together, locally: your material stays on your device, the model works from it, the relationships are explicit, and a change anywhere can be traced to what it affects. Each of these is explained from the ground up in the concept map.
Why this is still not an "AI researcher"
Stack all of it (a private local model, given tools, coordinated over a graph with memory, RAG, and GraphRAG) and you have a capable, private, inspectable assistant. You still do not have something that can be the researcher.
Every layer here decides where to look and what to show you, faster and more privately than before. Not one of them decides what is true or what is sound — that judgment is made by a qualified person about a specific claim, in light of specific evidence and assumptions.
The one line to keep. Local, agentic, and orchestrated change where the work runs and how well it is tracked, not who decides. Done honestly, this stack does not replace the researcher's judgment; it makes that judgment better-documented and easier to defend.
See the local, human-first approach in the tools.
Vahtian's research tools are built on exactly this division of labour: run on your own device, keep the human deciding, and record the trail. StudyVahti helps you structure and document a study before data collection; QualiVahti Local is the packaged local qualitative workflow: transcription, coding under your review, a codebook, evidence maps, and an audit trail, on your own computer.
Related guides
Start with the AI-concepts map for the whole stack in one table. Then see can a local model check your citations? for an honest laptop-model test, is it ethical to use AI on interview data? for the privacy case, and AI qualitative coding, human-reviewed for the division of labour in practice. More guides are on the Learn page.
One honest letter a month
One research problem, what I tried, what worked, and where it still falls short. Unsubscribe in one click.
Double opt-in: nothing arrives until you confirm by email. No tracking pixels, never shared. Privacy