← redevops.ioAll reference implementations →
Reference implementation · LlamaIndex

LlamaIndex retrieves.
ReDevOps completes the closure.

A real llama-index-core v0.14.24 VectorStoreIndex retriever and query engine run in-process with the Mission SDK, Closure Resolution and governance — real indexing and retrieval on a local offline embedding. No mock, no sidecar approximation.

Evidence · controlled · internally reproduced 27 Oct 2026 on the pinned code: closure recall 0.320 to 0.693 and contract-caller 0.300 to 0.600, identical.

LlamaIndex
VectorStoreIndex · retriever
query engine · Workflows
indexing & embeddings
the retriever — kept intact
→
rAIls
Context / Closure
Authority · Replay
Verification · Governance
the retriever as one representation, composed

LlamaIndex's strongest abstraction is retrieval, so we test the boundary immediately outside it. A semantic retriever finds what is similar; it does not, by construction, recover a code change's structural closure, and its query engine synthesizes a confident answer whether or not retrieval found the needle. ReDevOps registers the LlamaIndex retriever as one representation and composes it with structural closure + evidence verification — composed and preferred when it suffices, never replaced.

Code closure · agentic-os

Similar isn't complete

0.320 → 0.693

Over 1,138 real symbols the retriever recovers the seed and type edges but misses the contract callers a correct change must touch (structural, not similar). Composing it with the call-graph closure roughly doubles recall — contract-caller recall 0.300 → 0.600 — with 0 regressions.

Long-context · query engine

Synthesis is confident; verification is honest

6/6 abstained

On a needle-in-a-haystack, when retrieval misses the needle the query engine still answers confidently (4 of 6). ReDevOps evidence-support verification abstains on all 6 unsupported answers — selective accuracy 1.000 @ 50% coverage.

Governed action

Behaviour, not recall

0 unauthorized · 0 duplicates

Retrieval informs; ReDevOps governs. A retrieved-evidence change runs only through a Mission: approval held; exactly-once execution; replay reconstructs from the event ledger and fires zero extra side effects.

The flagship result is real and modest: retrieval already resolves the closure on the easy tasks, and composition never degrades those. The composed arm is structurally seeded, so its match recall is high by construction — that is not claimed as a retrieval win. The meaningful deltas are closure and contract-caller recall.

View the LlamaIndex reference implementation → Run the benchmark → Read the architecture article →
// conformance summary   every native-vs-native+ReDevOps comparison is classified — PARITY / FRAMEWORK_NATIVE_ADVANTAGE / REDEVOPS_DELTA / EXPECTED_IMPLEMENTATION_DIFFERENCE / BUG. LlamaIndex suite: 10 REDEVOPS_DELTA · 1 PARITY · 1 EXPECTED_IMPLEMENTATION_DIFFERENCE · 0 BUG. Full tables, environment and acceptance hashes on the benchmark page.

// llama-index-core 0.14.24 · Mission SDK 0.2.0 · Context Runtime 0.2.0 · one Python env · github.com/redevops-io/redevops-integrations

LlamaIndex is a trademark of LlamaIndex, Inc. This is an independent integration, not a LlamaIndex product or endorsement.