Saltar al contenido principal

pix

Since v1.14.0 · Top-level declaration

Grammar​

pix <Name> {
    source: "<uri-or-name>"       # required — the document/corpus the PIX indexes
    depth: <integer>              # optional — max navigation depth d_max (1..8)
    branching: <integer>         # optional — max branching factor b_max (1..10)
    model: <ident>               # optional — the navigation LLM slug (e.g. fast)
}

pix (Progressive Index eXtraction) declares an embeddings-free structural retrieval index: a navigable tree over a document or corpus that an LLM traverses with intention, mirroring how a domain expert consults a complex document — skim titles, drill into promising sections, follow cross-references — instead of nearest-neighbour search over chunk embeddings.

PIX replaces RAG's axiom, it does not implement it. RAG: Relevant(chunk, q) ⟺ cos(φ(chunk), φ(q)) ≥ θ. PIX: Relevant(section, q) ⟺ I(R; section | q, path) > ε — a section is relevant iff visiting it reduces uncertainty about the answer, given the query and the navigational path already taken.

There is no embedding pipeline, no vector database, and no cosine similarity anywhere in a PIX. The formal basis is papers/paper_pix_formal_research.md (the Chunking Destruction Theorem; monotone conditional-entropy reduction; reasoning-path explainability).

Note — naming (v2.12.0). pix was briefly documented as a "Provenance Index" (an audit chain). That role now lives in the ledger primitive. pix is, and always was in the grammar (navigate <pix> takes a pix), the retrieval navigator described here.

Surface​

pix is a top-level declaration. The navigation verbs navigate, drill, and trail operate over a declared pix.

pix ContractIndex {
    source:    "contracts/master_agreement.pdf"
    depth:     4
    branching: 3
    model:     fast
}

flow AnalyzeClause(question: String) -> Analysis {
    navigate ContractIndex {
        query:  "${question}"
        trail:  true          # record the reasoning path
        output: sections
    }
    step Analyze {
        given: sections
        ask:   "${question}"
        output: Analysis
    }
}

Fields​

source: (required)​

A string literal naming the document or corpus the PIX indexes — a file/URI to index into a tree, or the name of a declared corpus. The runtime builds (or loads) the document tree D = (N, E, ρ, κ): each node carries a ⟨title, summary, location, children⟩ representation whose summary is an intentionally lossy compression (target ratio 5–15%), sufficient to decide whether to explore deeper, not to answer.

depth: (optional)​

A non-negative integer in 1..=8 — the maximum navigation depth d_max. Bounds the traversal (Theorem: navigation converges in at most d_max steps). Default: a small depth suited to the document's height.

branching: (optional)​

A non-negative integer in 1..=10 — the maximum branching factor b_max, i.e. how many ε-informative children the navigator expands per level after pruning. The greedy expansion is (1 − 1/e)-optimal by submodularity of the information-value function.

model: (optional)​

A single identifier naming the navigation LLM (e.g. fast). The navigator uses a lightweight model to score f_LLM(q, node.summary) ≈ I(R; node | q, path); the final answer is generated separately by the flow's reasoning step. This tiered split (cheap navigation, powerful generation) is the cost model in the paper.

Epistemic integration​

PIX output rides the epistemic lattice and is never auto-trusted:

node summary (intermediate)  →  speculate   (lossy, navigation-only)
leaf content (terminal)      →  believe     (real content, external provenance)
validated content            →  know        (only after anchor/shield validation)

Effect row: ⟨io, epistemic:believe⟩ — reads controlled, pre-indexed content; trusted but unverified until validated.

What this primitive is NOT​

  • Not a vector store. No embeddings, no pgvector/chroma/qdrant. Retrieval is guided tree traversal, not nearest-neighbour search. For the embeddings-backed surface see corpus.
  • Not an audit chain. The append-only, hash-linked tamper-evident record is ledger.
  • Not per-corpus. A single pix indexes one document/corpus tree. Cross-document graph navigation (MDN) composes multiple sources.

See also​

  • axon://primitives/corpus — the multi-document corpus a PIX can index.
  • axon://primitives/ledger — the audit chain (former Provenance-Index reading of pix).
  • papers/paper_pix_formal_research.md — the formal framework.
  • papers/paper_multi_document.md — MDN, cross-corpus navigation.