SearchRequest.radius is declared, range-validated to [0.0, 1.0], and documented at memory/search/dto.py:84-88 as gating "raw cosine at recall time". It has no effect anywhere.
- HYBRID:
manager.py passes radius=_effective_radius(req) into RankInput at three call sites (~:420, :477, :551). grep -rn radius .venv/.../everalgo/ returns exactly two hits, both in types/rank.py — the field declaration and its docstring. everalgo never reads it.
- AGENTIC: never passed at all.
PR #393 briefly wired it into the agentic path as ahybrid_retrieve(min_score=radius) and had to revert: min_score is applied after RRF fusion (everalgo/rank/hybrid.py:86-91), where rrf replaces each score with Σ 1/(k+rank) (max ≈ 0.0328, fusion.py:41). Since top_k defaults to -1 and _effective_radius then returns _DEFAULT_UNLIMITED_RADIUS = 0.5 (manager.py:120, :937-941), every candidate was filtered out — default agentic agent search returned []. Measured: {None: 2 hits, 0.5: 0 hits, 0.03: 2 hits}.
Needs a decision before any implementation: what should radius mean? Candidates — (a) a pre-fusion cosine floor applied inside the dense recall route, which is what the docstring describes; (b) a post-fusion floor on the calibrated LR score, which is what min_score already does; (c) remove it and keep min_score. Whichever is chosen, apply it consistently across lanes and delete the three inert RankInput.radius args.
Found while verifying #393.
SearchRequest.radiusis declared, range-validated to[0.0, 1.0], and documented atmemory/search/dto.py:84-88as gating "raw cosine at recall time". It has no effect anywhere.manager.pypassesradius=_effective_radius(req)intoRankInputat three call sites (~:420,:477,:551).grep -rn radius .venv/.../everalgo/returns exactly two hits, both intypes/rank.py— the field declaration and its docstring. everalgo never reads it.PR #393 briefly wired it into the agentic path as
ahybrid_retrieve(min_score=radius)and had to revert:min_scoreis applied after RRF fusion (everalgo/rank/hybrid.py:86-91), whererrfreplaces each score withΣ 1/(k+rank)(max ≈ 0.0328,fusion.py:41). Sincetop_kdefaults to-1and_effective_radiusthen returns_DEFAULT_UNLIMITED_RADIUS = 0.5(manager.py:120,:937-941), every candidate was filtered out — default agentic agent search returned[]. Measured:{None: 2 hits, 0.5: 0 hits, 0.03: 2 hits}.Needs a decision before any implementation: what should
radiusmean? Candidates — (a) a pre-fusion cosine floor applied inside the dense recall route, which is what the docstring describes; (b) a post-fusion floor on the calibrated LR score, which is whatmin_scorealready does; (c) remove it and keepmin_score. Whichever is chosen, apply it consistently across lanes and delete the three inertRankInput.radiusargs.Found while verifying #393.