epistemic-vectors 0.1.0

Epistemic Vectors — a provenance-native primitive for retrieval pipelines. Drop-in Epistemic Layer for LangChain, LlamaIndex, and Mem0.

epistemic-vectors

A provenance-native primitive for retrieval pipelines — the Epistemic Layer drops in between vector-store retrieval and generation in any RAG stack and attaches an Epistemic Vector to every returned document.

What semantic embeddings did for meaning, Epistemic Vectors do for belief state. Five typed slots per document — revision_state, provenance_depth, consistency_score, authority, freshness — derived deterministically from a belief graph. No LLM, no free-text classifier.

Install

pip install epistemic-vectors[langchain]       # LangChain adapter
pip install epistemic-vectors[llamaindex]      # LlamaIndex adapter
pip install epistemic-vectors[mem0]            # Mem0 adapter
pip install epistemic-vectors[all]             # all three

Quickstart — LangChain

from agm_r import KnowledgeGraph, assert_belief, revise
from epistemic_layer.adapters.langchain import EpistemicRetriever

# (1) Build a belief graph
kg = KnowledgeGraph()
kg = assert_belief(kg, term="pluto_classification", source_id="doc_a",
                       claim_text="Pluto is the ninth planet",
                       timestamp="1930-02-18")
kg = revise(kg, old_source_id="doc_a", new_source_id="doc_b",
                new_claim_text="Pluto is a dwarf planet",
                timestamp="2006-08-24")

# (2) Wrap any LangChain retriever (e.g. Chroma) with the Epistemic Layer
retriever = EpistemicRetriever(base_retriever=chroma_retriever, kg=kg)

# (3) Docs now carry an Epistemic Vector in their metadata
for d in retriever.get_relevant_documents("what is Pluto?"):
    print(d.metadata["epistemic_vector"])
    # {"source_id": "doc_b", "revision_state": "asserted", ...}

Filter or re-rank by vector slots:

retriever = EpistemicRetriever(
    base_retriever=chroma_retriever, kg=kg,
    filter_by=lambda ev: ev.revision_state != "retracted",
    rerank_by=lambda ev: 0.5 * ev.freshness + 0.5 * ev.consistency_score,
)

Quickstart — LlamaIndex

from epistemic_layer.adapters.llamaindex import EpistemicRetriever
retriever = EpistemicRetriever(base_retriever=vector_retriever, kg=kg)
nodes = retriever.retrieve("what is Pluto?")
for nws in nodes:
    ev = nws.node.metadata["epistemic_vector"]

Quickstart — Mem0

from mem0 import Memory
from epistemic_layer.adapters.mem0 import EpistemicMemory

base = Memory()
em = EpistemicMemory(base=base, kg=kg)
em.add(messages="Pluto is a dwarf planet", user_id="alice",
       metadata={"source_id": "doc_b"})
results = em.search(query="what is Pluto?", user_id="alice")
for r in results["results"]:
    print(r["epistemic_vector"])

The primitive itself (no adapter)

Use epistemic_vector() directly against any agm_r.KnowledgeGraph:

from agm_r import KnowledgeGraph, epistemic_vector
ev = epistemic_vector(kg, "doc_b")
print(ev)
# EV(doc_b, asserted, prov=0, cons=1.00, auth=1.00, fresh=1.00)

Links

• Paper: Epistemic Vectors and the Epistemic Layer: A Provenance-Native Primitive for Retrieval Pipelines — NeurIPS 2026 (in submission)
• AGM-R (Python reference implementation): agm_r/
• License: MIT