1What was assessed
~283,000 lines of source (155 Python scripts and 396 UI/UX scripts), six ranked subsystems, 111 live technical documents and the integration patterns between them. The add-ons and learning companions are covered in C5.
6
Ranked subsystems
~283K
Lines of code
111
Live technical documents
3
Platforms: web, iOS, Android
2Scores by dimension
| Dimension | Score |
|---|---|
| Documentation completeness | 9.5 |
| Innovation & features | 9.2 |
| Overall architecture & design | 8.9 |
| Subsystem integration | 8.7 |
| Security & governance | 8.5 |
| Operational readiness | 8.5 |
| Code quality & testing | 8.0 |
Code quality and testing is the lowest score. This is where our technical team should look hardest; see "what to probe" below.
3Key findings
Architecture
- One corpus, many products. A single knowledge base grounds the chatbot, Content Studio, the LMS and the API, instead of several disagreeing copies.
- Modular. The six subsystems are cleanly separated and independently deployable; integration happens through defined interfaces, without tight coupling.
- Multi-tenant by design. One instance per company, each with its own bot, LMS, API and isolated MySQL database, over a per-company RAG layer.
- Portable. An export bundle packages platform code, the instance directory, a MySQL dump and a RAG export into one file.
Security and operations
- Credentials load from a database config table, not hardcoded or in environment files.
- Bearer-token authentication and per-key rate limits on API surfaces.
- Fail-explicit doctrine: no silent error catching and no default fallbacks.
- Structured logging with daily rotation, health probes per service, retrieval logs, conversation audit.
- No vendor APM agent is bundled, deliberately: monitoring tooling is left as our infrastructure decision.
Technology stack
| Retrieval | PostgreSQL + pgvector, vector(1536), HNSW index (m=16, ef_construction=64), retrieve-then-rerank |
| Models | Embedding: Cohere embed-v4.0 · Rerank: Cohere rerank-v3.5 · Generation: Claude, GPT, Cohere, self-hosted Llama |
| Ingestion | 40 file formats, 20+ loaders, OCR fallback for scanned PDFs, audio/video transcription |
| Delivery | 258 REST endpoints, MCP server, tool-using agent, streaming chat, report generation |
| Interface | UI/UX scripts for the administration and application screens; Flutter mobile apps |
4Objections the review answers
| Objection | Answer in the review |
|---|---|
| "This locks us into a vendor." | Full source is delivered, uncompiled. LLM-agnostic, on standard PostgreSQL/MySQL, open protocols, no license server. |
| "We do not have an AI team." | We need an AI Data Expert and engineers, not researchers. The specialist choices (embeddings, chunking, indexes) are made and documented. What remains is ordinary application engineering: REST, queues, auth, front ends. |
| "Why not LangChain or LlamaIndex?" | Every pipeline stage is a readable file, so the receiving team can learn and modify it, rather than working behind framework abstractions. |
| "Can we use our own front end?" | Yes. All 258 REST endpoints can be consumed from any front-end language without touching the shipped UI/UX scripts. |
| "There is no observability story." | Structured logging, per-service health probes, per-tool usage counters, retrieval logs and a conversation audit surface. The APM tool is left to us. |
| "Where are the benchmarks?" | It ships the instrument instead of a score: a QA workflow, per-answer citations and a debug payload, to evaluate on our own corpus. |
5What our team should probe
| Probe | Our finding |
|---|---|
| Automated test coverage and the regression process (the lowest-scored dimension, 8.0) | [p_c1_1] |
| UI/UX scripts skills: do we have them, or will the certified integrator cover that layer? | [p_c1_2] |
| Disaster-recovery procedure on our infrastructure | [p_c1_3] |
| Fit with our monitoring / APM tooling | [p_c1_4] |
6Verdict, as stated by the review
Approved for adoption, grade A-. A mature, coherent, production-ready system. Its defining advantage is coherence: engine, interface, knowledge, documentation and training were designed as one system. Success depends on a mandatory AI Data Expert on the customer side, plus a certified system integrator or the customer's own IT team, and, for mission-critical or compliance-heavy use, an organization-specific security and performance review.
Our assessment: [as_c1]. [as_c1_why]