fix: clamp inf/nan similarity scores at source to prevent JSON crash

[snemesh]

Problem

Closes #871

The C++ vector engine (bruteforce.h) can produce inf float values when computing inner product similarity on non-normalized vectors (distance_type=ip with NormalizeVector=False). The raw dot product of high-dimensional vectors with large component values overflows float32 to +inf.

These inf scores propagate through the Python adapter and retriever layers untouched, until JSON serialization fails:

ValueError: Out of range float values are not JSON compliant: inf

This crashes the entire server process, requiring a manual restart.

Root Cause

1. CollectionAdapter.query() reads item.score from the C++ engine result and assigns it directly to record["_score"] — no finite check
2. HierarchicalRetriever._recursive_search() reads _score and applies score propagation: alpha * inf + (1 - alpha) * current_score = inf — amplifying the problem
3. The existing isfinite guard in _convert_to_matched_contexts() catches scores only at the final conversion step — too late, because inf already reaches JSON serialization in intermediate API responses

Fix

Clamp non-finite scores to 0.0 at the two layers where they enter the Python pipeline:

Layer 1 — openviking/storage/vectordb_adapters/base.py

• CollectionAdapter.query(): check math.isfinite() immediately after reading from engine result, before the score enters any JSON-serializable record

Layer 2 — openviking/retrieve/hierarchical_retriever.py

• Clamp _score at all three entry points where scores are read from search results:
  • _merge_starting_points() — global search scores
  • _prepare_initial_candidates() — initial candidate scores
  • _recursive_search() — per-directory search scores

This ensures inf values never propagate into score arithmetic or JSON responses.

Why not PR #880?

PR #880 applies SafeJSONResponse only to error handlers. This prevents the server crash, but the search endpoint still returns an HTTP 500 error instead of results — the user gets nothing. This PR fixes the actual scores so search returns correct results.

Changes

• openviking/storage/vectordb_adapters/base.py — 1 import, 3 lines changed
• openviking/retrieve/hierarchical_retriever.py — 9 lines changed (3 locations)

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Failed to generate code suggestions for PR