Teams can search and question confidential documents without sending them to an external model API.
The system searches and prepares answers on client-controlled infrastructure. Users receive an answer with the source document and passage. Operators manage access and classification, while the environment blocks connections to external AI services.
Problem, solution, and result
Why a standard cloud assistant is unsuitable
Confidential documents cannot leave the company network. Security needs a technical way to test the boundary, and users need to see which document supports an answer.
What the system does
It indexes 180 documents inside the environment, retrieves relevant passages, and prepares answers locally. Every answer cites a source. Access and classification are checked before content is returned.
What we tested
The demonstration confirms that documents and answers remain inside an isolated environment and that outbound connection attempts are blocked. It uses 180 synthetic documents. Search quality, answer accuracy, and performance must be confirmed on the client’s documents and hardware.
For whom
This fits an organization that needs search across confidential documents, requires local processing, and wants a repeatable outbound-traffic test.
Documents → retrieval → local model → source
- 01Document content stays inside the company network
- 02Every answer cites a document and passage
- 03Security can repeat the outbound test
- Best for
- Regulated organizations that must process confidential documents locally
- Input data
- The team's questions against a local corpus of confidential documents
- Human decision
- Operators manage access and classification, while security approves the environment
- Cost
- PLN 0 in external API fees; 3 answers prepared locally on GPU, 430 ms median and 529 ms at the 95th percentile.
- Test scope
- The demonstration checks that no connection leaves the environment. The client’s security team repeats this check in its own infrastructure.
- Audit trail
- The history shows every classification change with the operator, previous value, and time.
- Similar system build
- from €6,000 net (approx. $6,500) · 4–10 weeks
Network and access boundary
Who owns access and results
Operators grant access, change classifications, and resolve conflicts. Users check answers against cited documents. Security approves the environment and repeats the boundary test.
- Cost
- PLN 0 in external API fees; 3 answers prepared locally on GPU, 430 ms median and 529 ms at the 95th percentile.
Estimated impact
Calculate the impact at your volume
This estimate uses the stated volume. Enter your own numbers to assess the possible impact in your company. The result is confirmed only during a pilot.
Today
167 h
With the system
58 h
Time or cost saved
80-135 h/mo, base 108 h
- Volume
- 1,000 queries/mo
- Formula
- 1,000 × (10.0 - 3.5) min / 60 = 108 h
- Calculation status
- low-medium
Data in the screenshots. The names, amounts, and documents shown in the screenshots are synthetic. Client data remains private. Measurements describe the demonstrator and are kept separate from production outcomes.
Working views
Operators manage access. Users verify sources. Security controls the network boundary.
The demonstration uses 180 synthetic documents. Deployment requires measuring search quality, answer accuracy, file-format coverage, and performance on the client’s hardware.
Document library, sourced answer, and boundary status.
Document library
Documents, classifications, and access history.
Sourced answer
The answer with the document and passage used to prepare it.
Environment status
Local dependencies and outbound-test results.
System screens
See how the system works in practice
These desktop and mobile screenshots come from the running application. They show the process and the points where a person makes the decision.
- Screens
- 10
- px
- 1440 · 390
- 021440×1353
Answer linked to its source passage. - 031440×1100
Outbound-traffic test result.
Open the remaining screen archive (7)
- 041440×1100
Reading room, answer with sources - 051440×1100
Reading room, material outside the index - 061440×1100
Reading room, no matching source - 071440×1100
Registry journal - 081440×1100
Markings to clarify - 091440×1100
Protected zone - 101440×1100
Test protocols
Technology stack
Open components next to the data, because the documents cannot leave.
This process gets a local stack because its one hard requirement is that content never leaves the network. A Polish open-weight model on vLLM replaces the cloud API, local embeddings with a reranker replace hosted search, and pgvector keeps the index in the same database as the documents. For an audit this means the boundary is checked with a network test rather than a data processing agreement, and what remains to maintain are standard open-source components.
- vLLM + Bielik 11B
- the target GPU profile, serving a Polish open-weight model from a read-only directory
- Local embeddings and reranker
- multilingual MiniLM with a cross-encoder runs retrieval without any hosted API
- PostgreSQL 17 + pgvector
- holds the documents, the vector index, and the change trace in one local database
- FastAPI
- runs the workflow, the clearance checks, and the explicit denial states
- Bielik 4.5B on MLX
- the recorded host-side proof on an Apple M5 Pro, measured separately from the vLLM profile
The client takes over the code, prompts, evaluation data, and documentation; everything runs on their infrastructure and touches no external AI service. We use open-source components their team can maintain without us.
Technical details and measurements
How the system works
The entire path from question to answer stays with the documents.
Local models retrieve passages, rank them, and generate the answer. The system cites the source and distinguishes no match, access denial, and file errors.
Confidential documents remain in the company network
Local retrieval selects relevant passages
Bielik prepares an answer with a source
The test probes HTTP, HTTPS, DNS, and TCP
The user verifies the answer in the document
System architecture
Local models, an internal layer, and a probe.
Embeddings and a reranker on site, generation in a layer with no exit, and a controlled four-protocol test.
- 01
Retrieval
Search on local models.
Multilingual embeddings and a cross-encoder run inside the stack, and the vector index lives in pgvector next to the documents. No query leaves the environment, because there is no hosted API for it to reach.
- 02
Generation
Bielik in the internal layer.
The generative model is reachable only from the stack's internal network, served by vLLM from a read-only directory. A clearly labeled degraded profile exists for GPU-less review; no published number comes from it.
- 03
States
No result is an explicit state.
No match, an unavailable index, an unsupported file, and a permission denial look different on screen and lead to different next steps. The system never covers ignorance with a vague answer.
- 04
Probe
Four protocols, zero connections.
A controlled test tries to leave the model layer over HTTP, HTTPS, DNS, and raw TCP. The result lands in the evidence and can be rerun on demand; four attempts do not replace certification of a client environment.
The exit boundary as a command you can run
The declaration that no data is sent is a sentence in an offer. The egress test is a command you can run on a Wednesday at ten with the auditor watching. The credibility of an on-prem system rests on that difference, which is why the stack stays as small as it can: every extra component is another outbound path to check, so every component creates another boundary to verify.
- No external model SDKs anywhere in the stack
- The model layer lives on an internal network
- The egress test can be rerun on demand
- Classification changes carry an operator and a history
Want to assess a similar process in your company?
- A 30-minute call with the engineer who would lead the work.
- A review of the processes that cost you the most time and money.
- A written summary of what to automate first and the likely cost range.
Times are shown in your own time zone. We work with clients across time zones.
The scan is free and creates no obligation. If automation is unlikely to pay off, the written recommendation will say so.