Cognosa Platform Overview

How a question is answered, and what the platform is made of

Cognosa lets an organisation ask questions of its own documents and get answers with sources, using whichever language model it prefers. Each client group has its own knowledge bases, users, instructions, and model choices, kept separate from every other group. The two figures below are the executive view; the detailed engineering diagrams sit behind them.

Figure 1. How a question is answered

back end · runs on our server private or cloud User Web app in the browser API server sign-in, records the request RAG engine background processor Knowledge base vector database Language model private or cloud asks a question, picks a knowledge base and a model 1 submits the question (signed-in user, their group) 2 records the request; web app checks progress 3 picks up the request 4 finds the most relevant passages (filters, top matches) 5 passages with their sources 6 question + passages + the group's instructions 7 answer, streamed as it is written 8 saves progress, then the final answer and sources 9 shows the answer with sources 10
The web request finishes at step 3. The remaining steps run in the background on the server while the web app checks progress, so a slow model never ties up the website. Solid arrows are requests, dashed arrows are results coming back. The knowledge base is a vector database; the language model can be a private one on our own hardware or a cloud service.

Figure 2. What the platform is made of

People each group (tenant) sees only its own data End users ask questions, read answers and sources Group administrators manage their group's users, knowledge bases, retrieval settings Platform administrator manages groups, models, settings, audit log Web application Ask choose knowledge base and model, ask, watch the answer arrive Review answer with the passages it was based on; export Prompt templates each group's instructions to the model Administration users, knowledge bases, models, settings, status API server Sign-in and roles user, group admin, platform admin Group isolation every request scoped to the user's group Request queue and status records each question; reports progress Configuration and audit models, knowledge bases, settings, change log RAG engine built on LangChain · runs in the background, several requests at once 1 · Embed turn the question into a vector 2 · Retrieve most relevant passages, with metadata filters 3 · Compose group template + passages + question 4 · Generate language model writes the answer, streamed Knowledge and application data Document ingestion PDF, DOCX, PPTX, XLSX, TXT split into passages with metadata, one tool run per set Knowledge base vector database (Qdrant) one collection per knowledge base Application database PostgreSQL: users, groups, settings, requests, answers, audit log Language models Private model open-source model on our own GPU server (Ollama) Cloud model service OpenAI, Anthropic, Google, Alibaba, via their APIs Chosen per group any mix; the model is picked when asking Hosting One server per environment AWS EC2: dev, demo, prod Containers web + API, engine, databases Secure access HTTPS, certificates renewed automatically
Read top to bottom: who uses it, what they see, the services behind the screen, the engine that answers, the data it draws on, the models it can call, and where it all runs. Every layer is shared across client groups, while the data and settings inside it are separated by group.

Where to look next

The three engineering diagrams carry the next level of detail: the deployment topology (which servers and containers run where), the component stack (the code, packages, and libraries including LangChain), and the query sequence (every step and status of one request).