Enterprise AI Vendor Procurement Assistant [Estimate your model use]

A grounded multi-agent network built on neuro-san. It answers LLM-procurement questions ("which vendor gives us HIPAA + GDPR + on-prem under $70k per year?") by fanning the request out to specialist agents, retrieving every fact from deterministic coded tools over local data, validating each claim independently, and only then producing a scored recommendation.

No answer relies on model memory. Missing, stale, or conflicting evidence survives into the final answer instead of being smoothed over.

LLM: Mistral mistral-medium-2505 | UI: nsflow at localhost:4173 | Orchestration: neuro-san

Architecture

• The Frontman never states a fact itself: it extracts requirements, asks for missing mandatory info, and delegates to seven specialists, each owning one dataset and one tool.
• Compliance delegates a level deeper to DataPrivacy (GDPR, residency) and Regulatory (HIPAA, FedRAMP) sub-agents.
• The EvidenceValidator independently re-reads the datasets and marks every claim supported / refuted / unverifiable; the DecisionAgent scores only validated evidence.
• Agents are declared in HOCON (registries/vendor_procurement.hocon); Python is reserved for the deterministic tools (coded_tools/).

Quickstart

# Linux / macOS / WSL
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env          # [add your MISTRAL_API_KEY](https://console.mistral.ai/?profile_dialog=api-keys)
./run_studio.sh

# Windows (PowerShell)
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env        # add your MISTRAL_API_KEY
.\run_studio.ps1

Open http://localhost:4173 and pick vendor_procurement. Other entry points: ./run_cli.sh (terminal chat, no server) and ./run_server.sh (API only, :8080).

Always launch through the run scripts: they load .env and set the absolute AGENT_MANIFEST_FILE / AGENT_TOOL_PATH / PYTHONPATH the server needs.

Try these

• Should we choose OpenAI Enterprise or Anthropic?
• We need an LLM for a healthcare startup in Europe. Budget $70k/year, need HIPAA, GDPR, on-prem, function calling, latency under 500ms.

How it works

Tools return a uniform evidence envelope (value, source, freshness, deterministic confidence, warnings), never prose. Confidence is 100 x freshness_weight x field_coverage; the fit score is a fixed weighted sum (Security 40, Compliance 20, Performance 15, Deployment 10, Pricing 10, Licensing 5) with explicit tie-breaks. Sixteen tested edge cases cover unknown providers, aliases, quote-only pricing, conflicting benchmarks, stale data, partial certifications, malformed cells, missing datasets, impossible constraints, and ties.

Testing

python -m pytest tests/ -q 

A committed record of a full end-to-end run (all agents, validator, decision) is in logs/e2e_query_summary.md.

Extending

• New provider: add it to providers.json and a row/section per dataset. No code change.
• New dimension: add a dataset, a SpecialistTool subclass, an agent in the HOCON, and a risk-score weight.
• Different LLM: swap llm_config in the HOCON and install the matching langchain-* package.