Flow A vs Flow B

Every candidate that enters the Raw to Knowledge pipeline is either Flow A or Flow B. This classification is assigned during intake and determines the path the candidate takes before it reaches the validation queue. The distinction matters for validators because the two flows carry different trust profiles.

---

Flow A: Answer present in the source

Flow A candidates arise when both a question and an answer are present in the ingested transcript. During intake, the extraction stage identifies the question-answer pair, captures the relevant segments, and creates a candidate record that already contains answer text drawn directly from the conversation.

Flow A is the primary path. The answer was spoken by a human in context — typically a subject-matter expert responding to a customer question, or a technical architect clarifying a product capability during a discovery call. The provenance is direct: there is a source segment, a speaker, and a timestamp.

When a validator reviews a Flow A candidate, the source segment is displayed alongside the candidate. The validator's task is to assess whether the extracted answer is accurate, complete, and correctly typed — not to question whether an answer exists.

---

Flow B: No answer in the source

Flow B candidates arise when a question was raised in the conversation but was not answered during the session. This happens when a question is deferred ("I'll follow up on that"), when the answer was unknown at the time, or when the discussion moved on before the question was resolved.

Raw to Knowledge flags these as Flow B candidates during intake. A Flow B candidate record holds the question text and the source segment but has no answer text. To produce an answer, a work item must be submitted to the AGWMP (Answer Generation with Managed Provenance) pipeline.

The AGWMP pipeline retrieves relevant knowledge from internal sources — Confluence pages, Jira issues, and existing approved answers in the registry — ranks them by relevance using a four-tier scoring system, and generates a candidate answer. The generated answer is scored by the F2 scorer (β=2, gate threshold ≥ 0.78). If the score meets the gate, a GeneratedCandidate record is created and enters the validation queue.

---

The known-answer shortcut

Before the AGWMP pipeline generates anything, it checks the Weaviate vector index for an existing approved answer that semantically matches the incoming question. If a match is found with a similarity score ≥ 0.92, the existing approved answer is surfaced directly — no generation occurs.

This shortcut serves two purposes. First, it avoids generating redundant candidates for questions the registry has already answered. Second, it surfaces the validated answer to the requester immediately, without waiting for the generation-and-validation cycle to complete.

Note: The known-answer bypass returns the existing approved answer as the response to the work item. The Flow B candidate record remains in the system for audit purposes, but it is marked as resolved by the existing answer.

---

The no-content escalation

If the AGWMP pipeline cannot find at least two qualifying RAG sources for a Flow B question, it does not attempt generation. Instead, the work item transitions to ESCALATED status and is routed to the specialist queue.

This behavior reflects an important governance principle: generating an answer from insufficient evidence is more harmful than acknowledging that the knowledge gap is real. An ESCALATED Flow B work item is a signal that the question requires a human to source the answer manually — by consulting documentation, engaging engineering, or conducting a follow-up conversation.

The governance lead is responsible for managing the escalation queue and ensuring that each escalated item receives a human-sourced resolution.

---

Trust posture differences

Both Flow A and Flow B candidates must pass through the same 9-state validation machine before entering the registry. An approved Flow B answer carries the same registry status as an approved Flow A answer. However, validators should apply different scrutiny depending on the flow.

Trust dimension	Flow A	Flow B
Answer origin	Human-spoken in conversation	LLM-generated from internal sources
Source traceability	Direct: transcript segment + speaker	Indirect: generation sources + RAG context
Citation review	Verify segment accuracy	Examine all citations carefully; check that sources are current and authoritative
Fabrication risk	Low (answer was stated)	Present (LLM may synthesize incorrectly from sources)
Recommended scrutiny	Standard	Elevated; check each evidence link

Important: A Flow B answer is not inferior by definition — a well-generated, well-cited answer from internal documentation can be more authoritative than an off-the-cuff statement in a meeting. But the validator must do the work of verifying citations. If any evidence link is missing, outdated, or irrelevant to the generated answer text, the correct decision is REVISED or REJECTED — not APPROVED.

---

Path diagram

flowchart TD
    IN([Transcript Ingested])
    DET{Flow detection\nduring intake}

    subgraph flowA["Flow A"]
        FA_C[Candidate created\nwith answer text]
        FA_V[Validation Queue\nValidator reviews\nsource segment]
    end

    subgraph flowB["Flow B"]
        FB_C[Candidate created\nno answer text]
        FB_WI[Work Item submitted\nto AGWMP]
        KA{Weaviate\nsim ≥ 0.92?}
        BYPASS[Existing approved answer\nsurfaced · no generation]
        SOURCES{≥ 2 RAG\nsources found?}
        GEN[AGWMP generates\ncandidate answer\nF2 score ≥ 0.78]
        ESC[Work item → ESCALATED\nHuman must source answer]
        FB_V[Validation Queue\nValidator reviews\ncitations carefully]
    end

    REG[(Registry\nApproved Answer)]

    IN --> DET
    DET -->|Answer present| FA_C --> FA_V --> REG
    DET -->|No answer| FB_C --> FB_WI --> KA
    KA -->|Yes| BYPASS
    KA -->|No| SOURCES
    SOURCES -->|No| ESC
    SOURCES -->|Yes| GEN --> FB_V --> REG