Axis Bridge Solutions
← All capabilities
Implemented Maturity: intermediate

Retrieval & RAG Systems

Hybrid semantic search with reranking over owned vector infrastructure — and the judgment to know when retrieval is premature.

Executive summary

Retrieval built as infrastructure, not a framework demo: pgvector on owned hardware, multilingual embeddings with cross-encoder reranking, tier-filtered results — plus documented maturity ladders that start retrieval-free on purpose.

What this capability means

RAG is the most commoditized term in AI — and the most commonly botched system. The failure mode is always the same: retrieval bolted on before the data model is right, evaluated by vibes.

Axis Bridge builds retrieval the other way around. The document store comes first (owned, backed up, access-controlled), embeddings and rerankers are replaceable components around it, and — where a corpus is still small — v1 ships retrieval-free with the upgrade path written down. That restraint is itself the capability: knowing what the next rung is, and not pretending to be on it.

Maturity ladder

  1. 01

    Basic — Context stuffing

    ● demonstrated

    Relevant documents selected by plain SQL and packed into the prompt — deliberately chosen for v1 systems where corpus size doesn't yet justify vectors.

    Query → SQL selection → Prompt context → LLM
  2. 02

    Intermediate — Hybrid semantic retrieval

    ● demonstrated

    Dense embeddings plus reranking over chunked documents, with access-tier filtering and whole-document handback.

    Query → BGE-M3 embedding → pgvector search → Reranker → Tier filter → Results
  3. 03

    Advanced — Evaluated, agentic retrieval

    ○ next

    Retrieval quality measured by eval suites; agents choosing retrieval strategies and tools per query.

    Query → Strategy selection → Multi-source retrieval → Eval-scored results → Agent

Evidence

  • Hybrid search live behind an authenticated doorway — BGE-M3 embeddings (1024-dim) with bge-reranker-v2-m3 cross-encoder ordering, over PostgreSQL 18 + pgvector. me-inc-os-knowledge / engine (deployed on Coolify)
  • Chunks inherit document privacy tiers; search finds a chunk, the doorway decides whether the whole parent document may be returned. me-inc-os-knowledge / PLAN.md §5
  • A written retrieval maturity ladder for the concierge platform — retrieval-free v1 → token-budgeted window → ANN retrieval → agentic tools — chosen deliberately, not by omission. whatsapp-concierge / README "Reply (v1)"
  • Knowledge packs structured for retrieval priority (numeric prefix ordering within knowledge bases). llm-knowledge-base conventions