Read-path scaling and observability: bounded candidates, tenant-filter pushdown, segment reranking, live metrics - #21

Open

ly-wang19 wants to merge 25 commits into

mainfrom

claude/intelligent-hypatia-25de98

Commits on Aug 12, 2026

Add cross-instance memory portability and MCP HTTP auth
claude
committed
8decc8c

Commits on Aug 16, 2026

docs: record the branch-debt triage
claude
committed
8d2eb36
retrieve: bounded candidate retrieval, off by default, plus the harness that measures it
claude
committed
d8bc6cd
store: push the tenant filter into the vector index instead of scanning for it
claude
committed
34f7f75
retrieve: skip the entity scan when the query has no negation, and push key lookups into LanceDB
claude
committed
dec8dc6
docs: record how the uncommitted worktree work was preserved
claude
committed
225d56f
rerank: score long sessions by segment instead of by their opening quarter
claude
committed
e43d64f
graph: index entity name terms so anchoring costs the query, not the store
claude
committed
6689523
service: live latency, volume and token metrics behind an open /metrics
claude
committed
8e33c66
retrieve: split the read context into a cacheable half and a per-query half
claude
committed
038fdc2
server: per-tenant rate limiting and Idempotency-Key replay
claude
committed
af49599
server: runtime-issued API keys, hashed at rest, behind their own admin token
claude
committed
ba8100d
client: a Python SDK with zero runtime dependencies
claude
committed
1925142
server: count what the rate limiter, idempotency cache and key resolution actually do
claude
committed
57f6475
docs: document the operational surface added over the last rounds
claude
committed
da6065d
proxy: make the system prompt stable across a session, opt-in
claude
committed
0eed3fc
eval: tell a real benchmark gain from the answerer's noise
claude
committed
b676a8b
eval: measure how much the benchmark moves when nothing changes
claude
committed
062e02e
eval: classify wrong answers by failure mode, and fix the log pointer that misdirects reproduction
claude
committed
5bd513a
eval: show that the refusals were not starved of evidence
claude
committed
af8040d
eval: print the significance test next to the accuracy table
claude
committed
d3eda07
eval: the evidence was already retrieved — 73% of failures had it in full detail
claude
committed
2617938
eval: correct the diagnosis — retrieval finds the evidence, assembly shows half of it
claude
committed
c63d0fd

Commits on Aug 17, 2026

eval: the counting fix does not clear the floor, and the measurement says why
claude
committed
7be088c

Commits on Aug 21, 2026

paper: drop an unused import
claude
committed
6791789