Prolo — RFQ Matching Data Flow (Stages 5–6)

Journey of a quote after ingestion: from the ExtractedQuote in the pipeline store → matched to a customer RFQ → submitted to the core system. Renumbered locally — Stage 1 = matching (was 5), Stage 2 = submit (was 6). This is the logical flow; matching internals are in rfq-matching.md. Yellow = AI-bearing. 

flowchart TB
    DB[("Pipeline store
ExtractedQuote (orphan)")]
    RFQ[("RFQ read-model
synced from core + embeddings")]

    subgraph M["1 · RFQ MATCHING"]
        direction TB
        NARROW["SQL narrow
supplier · date · qty/total ranges"]
        KNN["embedding kNN (recall)
quote line ↔ RFQ line"]
        RERANK["cross-encoder rerank (precision)"]
        LLM["LLM disambiguate
bundling · units · confidence"]
        NARROW --> KNN --> RERANK --> LLM
    end

    P["2 · SUBMIT TO CORE
update system of record"]

    DB -->|"read new quote's lines (EXTRACTED)"| M
    RFQ -->|candidate RFQ lines| M
    M -->|"per-line matches (line↔rfq_line, evidence)
→ roll up status = MATCHED"| DB
    DB -->|read matched| P
    P -->|"submit / update (P2)"| API[(Internal procurement API)]
    API -.->|periodic sync| RFQ

    M -.no/ambiguous match.-> H
    P -.rejected.-> H
    H["HUMAN REVIEW
cross-cutting"]
    H -.corrected.-> P

    classDef ai fill:#fde047,stroke:#ca8a04,color:#000
    class KNN,RERANK,LLM ai

    %% node / zone colours
    style DB fill:#eef1f4,stroke:#57606a,stroke-width:1.5px,color:#1f2328
    style RFQ fill:#eef1f4,stroke:#57606a,stroke-width:1.5px,color:#1f2328
    style API fill:#eef1f4,stroke:#57606a,stroke-width:1.5px,color:#1f2328
    style NARROW fill:#eef4ff,stroke:#3b7ddd,stroke-width:1.5px,color:#1f2328
    style P fill:#eef4ff,stroke:#3b7ddd,stroke-width:1.5px,color:#1f2328
    style H fill:#fdeef0,stroke:#d1242f,stroke-width:1.5px,color:#1f2328
Stage / process step Yellow = AI-bearing (kNN · rerank · LLM) Store / external API (pipeline · RFQ · procurement) Human Review (cross-cutting) Solid = data flow · Dotted = escalation / sync

RFQ Matching — decomposed sub-pipeline (Stage 1, was 5)

Matching is not one LLM call — it's a retrieve-then-rerank funnel that turns a new quote's line items into durable line_item_matches: SQL narrow → kNN recall → cross-encoder rerank → LLM disambiguate → confidence → write. Embeddings are precomputed at Stage 4; the cheap rerank clears easy pairs, the LLM only handles the hard/bundled ones. Yellow = AI-bearing. Diamonds are decisions. 

flowchart TB
    IN["new ExtractedQuote
(lines + precomputed embeddings)"]
    RFQ[("RFQ read-model
(synced, embedded)")]
    NARROW["SQL NARROW
supplier · date · qty/total ranges"]
    CAND{"any candidate
RFQ lines?"}
    KNN["kNN RECALL (pgvector)
top-K RFQ lines per quote line"]
    RERANK["CROSS-ENCODER RERANK
reorder pairs by precision"]
    THRESH{"top pair
confidently clear?"}
    LLM["LLM DISAMBIGUATE (Bedrock)
bundling · units · quantity · rationale"]
    CONF{"match_confidence
≥ threshold?"}
    WRITE["WRITE line_item_matches
(line ↔ rfq_line, evidence)"]
    ROLL["ROLL UP quote → status = MATCHED
enqueue to submit (P2)"]
    H["HUMAN REVIEW"]

    IN --> NARROW
    RFQ -->|candidate lines| NARROW
    NARROW --> CAND
    CAND -- no --> H
    CAND -- yes --> KNN --> RERANK --> THRESH
    THRESH -- yes --> CONF
    THRESH -- "no (hard/bundled)" --> LLM --> CONF
    CONF -- yes --> WRITE --> ROLL
    CONF -- "no / ambiguous" --> H

    classDef ai fill:#fde047,stroke:#ca8a04,color:#000
    class KNN,RERANK,LLM ai

    %% node colours
    style IN fill:#f2f0ff,stroke:#8250df,stroke-width:1.5px,color:#1f2328
    style RFQ fill:#eef1f4,stroke:#57606a,stroke-width:1.5px,color:#1f2328
    style NARROW fill:#eef4ff,stroke:#3b7ddd,stroke-width:1.5px,color:#1f2328
    style WRITE fill:#eef4ff,stroke:#3b7ddd,stroke-width:1.5px,color:#1f2328
    style ROLL fill:#eef4ff,stroke:#3b7ddd,stroke-width:1.5px,color:#1f2328
    style CAND fill:#fff8e6,stroke:#bf8700,stroke-width:1.5px,color:#1f2328
    style THRESH fill:#fff8e6,stroke:#bf8700,stroke-width:1.5px,color:#1f2328
    style CONF fill:#fff8e6,stroke:#bf8700,stroke-width:1.5px,color:#1f2328
    style H fill:#fdeef0,stroke:#d1242f,stroke-width:1.5px,color:#1f2328

Matching (Stage 1, was 5)

The hard part: there's no RFQ identifier in the email. Two layers — SQL narrow (date window · supplier · SKUs · totals) to shrink the search space, then LLM / semantic disambiguate among the narrowed candidates to match fuzzy line-item descriptions that share no key → emits rfqId, matchConfidence, matchEvidence[].

Submission (Stage 2, was 6)

Reads the matched row and calls the internal procurement API to submit/update the SupplierQuoteObject. Crossing into a system you don't own, so idempotency, revisions, and API-down are handled here; on success the row flips to submitted.

Human Review

No / ambiguous match (Stage 1) → review queue, using stored provenance + matchEvidence. Rejected submission (Stage 2) → review, corrected, resubmitted.