My API authentication rule is: All endpoints except /health and /metrics require X-API-Key matching the API_KEY environment variable.
My corpus contains 24 synthetic documents: orders 5, refunds 5, support tickets 5, warehouse logs 5, supplier reports 4.
My duplicate-refund handling is: Report both $249.99 and $429.99, cite both documents, flag the conflict, confidence: low; never silently pick one.
My return-policy conflict rule is: RP-2019 30-day policy is superseded; RP-2023 45-day policy is current; do not blend them.
My Phoenix negative control is: No Phoenix warehouse exists in the corpus; expected response is no evidence found with zero citations.
My Dallas causation caveat is: Carrier backlog cleared within three days of the packaging supplier change; the two causes were not isolated and improvement must not be attributed to packaging alone.
My evaluation suite has this scope: Ten fixed questions; evaluates citation recall, required content and confidence, minimum multi-hop count, and metadata-filter compliance.
My offline test suite has this setup: 84 offline tests; deterministic hash embeddings and scripted LLM; no API key required.
My retrieval configuration is: TOP_K=5; MAX_HOPS=3; CHUNK_SIZE/CHUNK_OVERLAP=500/50; MEMORY_TURNS=4.
My production search upgrade order is: Cross-encoder reranking is first production upgrade; hybrid BM25 plus vector search is second; BM25 is appropriate for exact-match recall.
I built NovaCart Intelligence Layer as a retrieval system over NovaCart operational documents. Its aim is to answer cross-functional questions by chaining evidence across document types, cite source documents, and make clear what the evidence cannot establish. I designed it to refuse to fabricate rather than present unsupported answers.
My NovaCart-Intelligence-Layer repository is a Python project on @ud-ai's GitHub; it started 2026-07 and was last updated 2026-07.
The system answers cross-functional questions by chaining evidence across document types and returns an answer that cites every source document.
My anti-fabrication design reports conflicting figures instead of picking one, marks inferred causal links as inferred, and says “no evidence found” when the corpus does not cover a question.
For a question, the agent plans sub-questions, uses semantic search and metadata filtering, and follows the linked_ids written into each document to pull in referenced records.
In my Dallas example, the answer chains refund records to carrier delay and carton failures, then to supplier non-conformance on packaging lot MPC-2402-B, with causation inferred from date correlation rather than stated in the records.
Every response carries a reasoning_trace that shows each hop, the filters applied, the documents found, and which were reached by link traversal.
Retrieval loops back when more evidence is needed and is hard-capped at MAX_HOPS; the README does not specify the numeric value.
My stack is Python 3.13; FastAPI; LangGraph; Google Gemini 2.5 Flash; Gemini text-embedding-004; persistent embedded Chroma with cosine similarity; SQLite; and a static API key via X-API-Key.
My FastAPI endpoints are GET /health, POST /documents, and POST /query; interactive docs are at /docs and the minimal web UI is at /.
Ingestion uses loader.py to validate and normalize, chunker.py to recursively split and add metadata, and embedder.py to call text-embedding-004 with retry.
My LangGraph agent stages are plan, retrieve, assess_hop, and synthesize.
My seed corpus contains 24 documents; after seeding, the health endpoint should report docs_indexed: 24.
Docker is my recommended setup; local setup uses a Python virtual environment and requirements.txt; a Google AI Studio API key is required, and the README identifies the free tier.
cp .env.example .env # then edit .env and set GOOGLE_API_KEY
docker compose up --build -d
curl http://localhost:8000/health
# Seed the 24-document corpus into the index
docker compose exec api python scripts/seed_data.py
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # then edit .env and set GOOGLE_API_KEY
python scripts/seed_data.py
uvicorn app.main:app --reload
My architecture diagram source is docs/architecture.mmd, and the README gives a Mermaid CLI command to export a PNG.
The README points to DESIGN_DECISIONS.md for rationale and scale considerations; that document's contents are not included in the material I have here.
The minimal web UI is a static page with no build step. I can use it to ask questions with live-streamed reasoning hops, generate reports, and inspect citations and traces without using curl. The example is useful for illustrating evidence chaining, but the underlying records and additional evidence are not included here.