Publication scope: raw profiling captures, transcripts, server logs, and memory traces are retained locally and omitted from this public branch. output/profiling/ paths and named review artifacts identify local evidence, not downloadable files. Hashes preserve evidence identifiers; the raw captures are unavailable for independent inspection in this repository.
EdBot uses SGLang 0.5.21 + native MTP to serve Qwen3.8-Flash-Next on the
DGX Spark, using NVIDIA's pinned NVFP4 checkpoint. Its large N-gram table is
served from local NVMe storage so the model fits the Spark's unified memory.
Ollama serves only qwen3-embedding:8b for course-material retrieval.
See the Spark setup guide for prerequisites, first-time
installation, SSH access, verification, and troubleshooting.
On the Spark, from this repository:
bash scripts/install_spark.sh --download-models
# Start Ollama using the command printed by the installer, then:
python3 scripts/start_sglang.pyAllow 260 GiB free disk space for installation, weights, and runtime files.
The installer downloads weights when requested; startup compiles GPU kernels and
creates a fresh roughly 48 GiB N-gram table file. Startup can take several minutes.
The launcher removes only its own temporary table directory after normal shutdown.
Wait for the server to become ready before starting EdBot. The chat endpoint is
http://127.0.0.1:30000/v1; port 8000 remains available for Chainlit.
Chat model selection and performance settings are defined in one code file,
src/edbot/model_config.py, shared by the client and server launcher. The
orchestrator and both subagents inherit the same chat client. The embedding model
is also selected in that shared configuration.
Set MODEL_PRESET = "27b" there to select the retained Qwen3.8-27B + DSpark
configuration, which uses its separate SGLang 0.5.19 environment. Flash-Next uses
the native installer; --docker is supported only by the 27B preset.
Install dependencies (first time only):
uv sync --lockedCreate a .env file in the repo root with at least:
CHAINLIT_AUTH_SECRET=... # generate one with: uv run chainlit create-secret
EDBOT_TEST_ACCOUNTS=alice:pw1,bob:pw2 # test logins, as username:password pairsThen, from the repo root (Chainlit reads .env and .chainlit/config.toml from the current directory):
uv run chainlit run src/edbot/ui/chainlit_app.py --host 0.0.0.0 --port 8000Open http://localhost:8000 and log in with one of the EDBOT_TEST_ACCOUNTS. Each username is its own
student, with its own conversation.
The first startup can take a while: EdBot clones the SEBook and builds its embeddings before the UI is ready.
Profiling tools compare warm inference and complete agent turns; see the profiling procedure. No speedup is assumed until comparable measurements and quality checks have been run.
The terminal interface uses the same backend: uv run edbot --student dev-cli -v.
uv run python -m unittest discover -s tests -v
python3 scripts/start_sglang.py --dry-runThe tests exercise streaming, tool calls, and structured subagent output against a local HTTP fixture. Real model correctness and performance must also be checked on the Spark; see the setup guide.
The latest tutoring quality review records successful execution of all nine synthetic turns, but unresolved answer disclosure, technical accuracy, and assessment issues. The current configuration is a measured performance candidate, not an accepted tutoring-quality release.
The selected tutoring policy retains concise explanations, Socratic practice and hints, and open-ended replies with the observed quality limits reported. Study-assignment handling remains as it is; a restricted reviewed-content delivery path is not selected.