Evidence for this page

Back to the page

Maturity: Early use (supported line 0.2.x).

Own statement

Date: 14 November 2026

There is no Gaia-X certification and no GDPR certification. A HelixTest run is not a GA4GH certification.

Source code

You enter a search. You save a hit in the library. A question looks for similar saved texts. A model writes an answer from them.

Present in the code

Date: 9 November 2026

Present in the code, no end-to-end run. The search calls PubMed. The question needs saved texts that already have vectors. Other screens are not these steps.

Source code

Further functions are present in the code: a biological file registry up to 500 MB, analysis workflows, BLAST, notebooks, Zenodo upload, and an export for the Medical Informatics Initiative.

Present in the code

Date: 13 November 2026

Whether these functions pass in a run is not shown.

Source code Source code Source code

Saved texts sit in the database whose address the operator sets. Stored files sit in a directory whose path the operator sets. A literature search sends the query to a service outside the premises. With no usable cloud key, a summary calls a model on the same computer. With such a key, that call goes outside. The operator chooses the provider, the model and the address. One request can choose the cloud provider when the key is set. The literature search calls PubMed. The model on the same computer is mistral:7b by default and runs through Ollama. A cloud key means Anthropic, default claude-haiku-4-5. The database holds the papers table with pgvector. Text search runs paraphrase-multilingual-mpnet-base-v2 in the program. An extra corpus named Locus is off by default. The code default for the provider is auto: a usable Anthropic key selects Claude, otherwise Ollama. The example file sets Ollama.

Present in the code

Date: 9 November 2026

This is in the code, but a full run and a network recording are missing, so it is not shown that nothing leaves the building or that the text-search weights are already on the computer. Zenodo, a Ferrum forwarder and remote workflows run only when the operator switches them on, and locus rows are shared and not tied to one person.

Source code Source code Source code Source code Source code Source code

BioResearch Assistant needs Docker and Python 3.11 or newer. The guide names at least 8 GB of RAM and 10 GB of disk, and recommends 16 GB and 30 GB. The model classes follow the size of the machine. A GPU is optional.

Present in the code

Date: 14 November 2026

The Compose memory limits are Ollama 16 GB, the backend 4 GB, and the frontend 512 MB. The 8 GB, 16 GB, 10 GB, and 30 GB figures in the install guide are guide figures, not Compose limits.

Source code

The source tree includes a getting-started guide, an install guide and an operator handbook, including for a computer without internet. They describe how to set the BioResearch Assistant up and update it without Synaptic Four. They do not state support terms.

Present in the code

Date: 9 November 2026

Present in the source tree, not installed in this review. No support hours, and no promise that Synaptic Four runs the system. The team price is a separate decision.

Source code Source code Source code Source code Source code Source code

In production, sign-in through the operator's sign-in system is required. By default each person sees only their own entries. A mode in which everyone sees everything does not start in production. The operator can share entries with a group.

Present in the code

Date: 9 November 2026

Present in the code, no end-to-end login. Use without sign-in exists only for local, development or test, and a production start refuses that. An unknown mode falls closed to each person's own entries. De-pseudonymisation defaults to the person who made the mapping. This program does not issue Passports.

Source code Source code

The BioResearch Assistant is under the Apache licence 2.0. You may use it, change it and pass it on freely, including commercially.

Present in the code

Date: 14 November 2026

Not legal advice. The text of the licence file is what counts.

Source code