A folder becomes an index
Lens processes documents where they already are: a selected root, parsed and chunked in place. The library stays in the file system, and the index appears beside it.
Instancium Lens · Knowledge workspace
Lens reads a folder of documents and answers questions in natural language. Each answer carries numbered references to the exact passage behind it, so a statement can be checked against the original.
The same workspace scales from a personal folder to a department or company knowledge base held inside the corporate network. Access follows the organization's identity and policy, retrieval runs on compact models, and the knowledge base remains an asset of the company.
For the person who needs the answer
The workspace speaks in folders, questions and sources. Retrieval and model mechanics stay underneath; verification stays with the reader. A question is asked the way it would be asked to a colleague, and the answer arrives with the passages it came from.
Lens processes documents where they already are: a selected root, parsed and chunked in place. The library stays in the file system, and the index appears beside it.
Files and passages are counted as they are processed, and the run ends with a summary of what was included. The state of the corpus is readable at every moment.
Answers stream token by token while the current stage stays on screen: searching the documents, then composing the answer. The pipeline reports its own progress.
Each claim carries a number bound to a file and a passage position. The source panel shows the original text, so a statement can be checked against the document.
Answers are composed from the indexed documents, and the sources stay attached. When the corpus holds a partial answer, the source panel shows the passages that were actually retrieved, so the boundary of the evidence is visible in the interface.
For CIO and CTO
A knowledge base becomes valuable when it is shared, and stays acceptable when it is separated. Lens keeps both properties by changing where the index lives. The topology below switches between Local and Corporate: the application, the citations and the way questions are asked stay the same on both sides.
One specialist, one device
The index is created next to the documents in .instancium and stays on the machine. Inference runs on the same device or on a gateway inside the corporate network - the endpoint is a configuration switch.
Single device. Owner scope.
One team, one shared collection
A knowledge base server inside the corporate network holds the data, the Vectra collections and the index. Indexing runs there as a separate service; the team asks the same collection and reads the same agreed documents.
Access mechanism selected by the organization.
Several domains, separated access
Policy, engineering and contracts live in separate collections. Identity and policy bind each profile to one collection, while the original files keep their own permissions in the file share or DMS.
Identity, policy, collection binding, file rights.
Documents folder — the selected root on this device; plain text and Markdown.
Lens App — parses, chunks, retrieves and answers with numbered citations.
.instancium — the Vectra index, created next to the documents.
InstantAIGate — on this machine or inside the corporate network; the endpoint is a configuration switch.
One device. The index stays beside the documents, and the device owner is the reader.
Corporate users run the Lens app on Windows.
Access layer — scoped API tokens, Active Directory / LDAP, API gateway policies, or IdentityServer / OIDC. The organization selects the mechanism; it binds each profile to one collection.
Knowledge Base server — source data, Vectra collections and the .instancium index.
Indexing service — a separate process on that server. Client workstations read the shared index; writing belongs to the service.
Source storage — originals in the file share or DMS; a cited file opens under file permissions.
InstantAIGate — server or cluster; embeddings and answer tokens.
Two contours of access. The collection decides what retrieval can find. File permissions decide which cited original opens. Both are administered inside the corporate perimeter.
The corporate topology is delivered as an integration project: the knowledge base server, the indexing service and the access mechanism are configured inside the organization's network. The Lens application ships as a Windows workspace and stays identical across modes.
Local mode: documents, index and citations stay on one Windows device. The inference endpoint is configurable: the gateway on that machine, or a gateway inside the corporate network. Scroll or + / − zooms · drag pans · 0 resets · Esc closes.
Institutional memory
A knowledge base held on company infrastructure keeps its value through staff changes. The indexed corpus, the citations and the profile bindings remain with the organization, and the accumulated answers stay available to the next specialist.
Institutional memory
A collection held on company infrastructure keeps its value through staff changes. When a specialist leaves, the directory account is deactivated while the corpus, its indexed passages and its citations remain in place. A successor receives a profile bound to the same collection and starts from the knowledge the organization has already accumulated.
Directory account: deactivated · Collection: retained · Profile binding: reassigned
Continuity follows the deployment scale: a shared collection persists on company storage, while a local index stays with its device.
Onboarding becomes a binding operation: a new profile points at the collection the role already uses, and the first question returns answers grounded in documents the organization collected over years.
Access model
Separation is a property of the deployment. It resolves in the corporate perimeter: the identity a person presents, the policy applied to a collection, and the binding between a workspace profile and that collection. The mechanism is selected by the organization; the inference gateway serves embeddings and answer tokens.
| Role | What the role does | Where the boundary sits |
|---|---|---|
| Reader | Asks questions over the collection their profile is bound to and opens the cited passages. | Identity from the corporate directory or a scoped token; the binding decides the collection. |
| Contributor | Submits documents for indexing into the collection their team owns. | Write permission granted in the access layer; indexing itself runs on the server. |
| Knowledge administrator | Owns collections, index storage, retention and the profile-to-collection bindings. | Directory groups and storage permissions inside the corporate perimeter. |
| Gateway operator | Owns model capacity, gateway health and gateway API keys. A control-plane role, separate from document access. | A separate gateway administrative scope, held by the operating team. |
Model size · domain fit · cost
Retrieval carries the company's knowledge; the model composes an answer from the passages that were found. This division of work keeps the required model size modest, keeps the cost of a question predictable, and leaves room for adapting the model to the organization's own language.
Model size
In a retrieval-grounded workload the facts come from the corpus, so the model composes an answer from the passages that were found. The reference configuration retrieves with a compact embedding model at 1024 dimensions and returns five passages by default. A compact chat model then writes the answer. Capacity planning becomes a hardware decision, and the cost of a question follows infrastructure capacity instead of a per-token tariff.
Corpus: facts · Compact model: composition · Cost: hardware capacity
Steady-state load stays small: a question embeds one short text and composes an answer from a handful of passages. Re-indexing follows the pace at which documents change.
Laboratory capability
The laboratory adapts a compact model to company terminology, document patterns and answer format: a curated corpus, parameter-efficient fine-tuning, evaluation against a held-out question set, then conversion to GGUF and serving through the same gateway. Citations, visible stages and the workspace stay unchanged, and the adapted weights are versioned alongside the collection they serve.
Curated corpus: LoRA adapter · GGUF weights · InstantAIGate · Same citations
Adaptation requires a curated corpus and an evaluation set. Quality is measured on the organization's own questions before the weights enter production.
Elastic capacity
A large corpus is embedded once. For that window an accelerator can be rented: the initial index is built at speed, then queries and incremental updates run on ordinary hardware inside the perimeter. The expensive phase is bounded in time and separated from steady-state operation.
Initial build: rented accelerator · Steady state: own hardware
A rented accelerator places the indexing workload with the provider for that window. The choice follows the classification of the corpus and the terms of the contract.
The economics follow the architecture: facts live in the corpus, composition runs on a compact model, and heavy compute is concentrated in the initial indexing window. Steady-state operation runs on hardware the organization already owns.
Controlled data perimeter
Running a knowledge base on company infrastructure gives an architecture review four concrete answers: where documents are parsed, where the index is stored, who decides access, and which component produces the answer.
The trade-off. Local mode keeps everything on one machine, and the index serves that machine alone. Corporate mode gives the organization one knowledge base and adds servers, a network hop, an integration project and an operational duty: the access topology, storage, indexing and model capacity become things a team operates. The scale follows the scenario.
Local and self-hosted deployment reduces how much sensitive content leaves the network. Access control, hardened configuration and operational procedures remain part of the deployment. Results depend on the models, the configuration and the rest of the system.
Deployment shape
A knowledge base is an operational commitment. The list below names the parts of a deployment and who owns them, so a pilot can be scoped against the architecture.
Scope boundary. The Lens application ships as a Windows workspace. The knowledge base server, the indexing service, the access mechanism and domain adaptation of the model are delivered as an integration project inside the corporate perimeter.
A common pilot shape: one department, one collection, plain text and Markdown documents, and a named owner for the access bindings, storage and model capacity. The domain set grows after answers have been checked against their sources in real work.