Scope — is / is not

We promise “no ungrounded claim” — not “zero hallucination.” We won’t pretend the model never errs; we guarantee it never asserts what the evidence doesn’t support. That honesty is why the core stays small.

It is

• An evidence-first RAG library you build products on.
• A retrieval layer over your evidence — dense + lexical + structure, with a graph and a navigate-not-cite wiki on top.
• An ingest engine that resolves every hit to a cited Evidence Unit.
• Typed intake verbs beyond documents: rag.ingest(...) for artifacts, rag.code.ingest_from(folder | git) for source (.py / .go today), and rag.schema.ingest_from(file | doc) for SQL DDL and OpenAPI/JSON-Schema. Each enforces its own contract — code and schema raise unless you declared the graph (or community) signal.
• Model-agnostic — your embeddings, LLM, reranker, and vision.
• S3-native, with reproducible, auditable provenance.
• Polyglot at parity — Go, JavaScript, Python, one shared conformance suite.

It is not

• Not a model host — nothing bundled; every model is injected.
• Not a code-comprehension product. Source and schema are first-class artifacts here — rag.code / rag.schema ingest them as citable Evidence Units, and every graph edge carries an explicit confidence label rather than being asserted as fact. What lives elsewhere is the product built on that: repo chat, call-graph analytics, IDE surfaces.
• Not a memory / “brain” and not an end-user app — those are built on top of the library, in their own repos.
• Not a promise of zero hallucination — a promise of nothing ungrounded.

Where a feature lands

A capability ships in the core only if it both ingests an artifact or improves grounded retrieval/evaluation and can be held to “no ungrounded claim.” A UI, a workflow, or a domain app is a separate product that consumes CiteNexus through its public API. The core stays small on purpose — proven capabilities can always be pulled in later; scope is hard to un-ship.