6.9 KiB

Raw Blame History

Contextual retrieval

One-line description. Anthropic's preprocessing recipe for RAG: before chunks are embedded and indexed, an LLM (the contextualizer) reads the whole document together with each chunk and emits 50–100 tokens of situating context that gets prepended to the chunk. The contextualized chunk then feeds both a semantic index (embedding model → vector DB) and a lexical index (TF-IDF → BM25). At runtime the query hits both indices, results merge through rank fusion, and the top-K chunks go to the generative model. The diagram's job is to show the contextualizer as the distinctive new step and the dual-track preprocessing it feeds.

Default diagram type

Flowchart (poster style) with two stacked phases. The pattern has a name, a clear preprocessing / runtime split, a fan-out into parallel tracks, and a distinctive new step — that's poster flowchart territory. A flat linear chart would smear the two phases and hide what's new. Stack the phases vertically with eyebrow dividers; each phase reads left-to-right.

Alternate types:

Structural — subsystem containers side by side when contrasting contextual retrieval against plain RAG (two siblings, each a mini pipeline). See structural.md → "Rich interior for subsystem containers".
Preprocessing-only flowchart when the runtime story isn't needed — drop phase 2 and end at the two indices.

Palette

Three accent ramps plus gray, under the poster-flowchart 4-ramp exception (ramps encode role categories, not sequence):

c-gray — corpus, query, rank fusion, top-K chunks, response. Neutral data / IO.
c-purple — Claude in both its roles: contextualizer and generative model. One ramp for both anchors the "same Claude, two prompts" story without adding a fourth color.
c-teal — semantic track (embedding model + vector DB).
c-amber — lexical track (TF-IDF + BM25 index).

Do not color the contextualizer and generative model differently — doing so implies different models or different roles, but the whole point is the same Claude doing both jobs.

Sub-pattern

flowchart.md → Poster flowchart pattern (eyebrow-divided phases, ≤4 ramps for role categories) + Fan-out + aggregator (simple mode) applied twice: the contextualizer splits into two tracks that never reconverge in phase 1, and query + both indices converge at rank fusion in phase 2.

Mermaid reference

flowchart TB
    subgraph Preprocessing
        C[Corpus] -- chunks --> CTX[Contextualizer · Claude]
        CTX -- context + chunk --> EM[Embedding model]
        CTX -- context + chunk --> TF[TF-IDF]
        EM --> VDB[(Vector DB)]
        TF --> BM[(BM25 index)]
    end
    subgraph Runtime
        Q[User query] --> RF[Rank fusion]
        VDB --> RF
        BM --> RF
        RF --> TK[Top-K chunks] --> GM[Generative model · Claude] --> R[Response]
    end

Defining edges: CTX --> EM and CTX --> TF (the contextualized chunk goes to both tracks) plus VDB --> RF and BM --> RF (both indices feed fusion). Drop either pair and the diagram collapses into plain RAG or embedding-only retrieval.

Baoyu SVG plan

Two stacked phases with eyebrow labels and a thin horizontal divider between them.

viewBox: 0 0 680 540
Phase 1 eyebrow — Preprocessing · Runs once per corpus update at (40, 50), class eyebrow.

Phase 1 interior:

Corpus — c-gray, x=40 y=80 w=100 h=56, two-line (Corpus, Documents).
Contextualizer — c-purple, x=180 y=72 w=260 h=72, multi-line (Contextualizer, Claude, 50–100 tokens per chunk). Visibly the largest box — it's the pattern's signature step.
Embedding model — c-teal, x=140 y=180 w=160 h=48, single-line.
TF-IDF — c-amber, x=380 y=180 w=160 h=48, single-line.
Vector DB — c-teal, x=140 y=260 w=160 h=56, two-line (Vector DB, Semantic index).
BM25 index — c-amber, x=380 y=260 w=160 h=56, two-line (BM25 index, Lexical index).

Phase 1 arrows:

Corpus → Contextualizer: (140, 108) → (180, 108), label chunks at (160, 102).
Contextualizer → Embedding model: L-bend (260, 144) → (260, 160) → (220, 160) → (220, 180), label context + chunk at (170, 164) text-anchor="end".
Contextualizer → TF-IDF: L-bend (360, 144) → (360, 160) → (460, 160) → (460, 180), label context + chunk at (470, 164) text-anchor="start". (Both arrows labeled — the reader must see that both tracks receive the contextualized chunk.)
Embedding model → Vector DB: straight vertical (220, 228) → (220, 260).
TF-IDF → BM25 index: straight vertical (460, 228) → (460, 260).
Phase divider — dashed line x1=40 y1=340 x2=640 y2=340, class arr-alt.
Phase 2 eyebrow — Runtime · Per user query at (40, 362), class eyebrow.

Phase 2 interior (single horizontal row at y=400–456):

User query c-gray x=40 w=100, Rank fusion c-gray x=160 w=100, Top-K chunks c-gray x=280 w=100 (two-line with subtitle Top 20), Generative model c-purple x=400 w=140 (two-line with subtitle Claude), Response c-gray x=560 w=80. All y=400 h=56.

Phase 2 arrows (straight horizontal, 20px gaps between boxes at y=428): query→fusion, fusion→top-K, top-K→generator, generator→response.

Cross-phase arrows (indices into rank fusion):

Vector DB → Rank fusion: vertical drop (200, 316) → (200, 400) — lands inside rank fusion's top edge (x=160–260).
BM25 index → Rank fusion: L-bend (460, 316) → (460, 372) → (220, 372) → (220, 400). The 20px x-offset from the Vector DB arrow keeps the two inbound arrows from stacking.

Both cross-phase arrows are solid .arr — they're the main data flow, nothing alternate.

Legend (bottom, required — 3 accent ramps encode category):

[■] Claude (contextualizer + generator)    [■] Semantic track    [■] Lexical track

Place at y=510, centered at x=340.

Gotchas.

Both tracks must show they receive the contextualized chunk — label both outgoing arrows from the contextualizer. If only one is labeled, readers assume the other track still uses raw chunks.
Do not draw the contextualizer as a self-loop on the Corpus. It's a distinct LLM step that runs once per chunk with whole doc + chunk as input, conceptually closer to an orchestrator than an inline transform.
Keep rank fusion gray, not amber — it merges two tracks but it's a structural aggregator, not an accent role. Giving it amber visually absorbs it into the lexical track.

Reranker variant. The reranking extension inserts a reranker box between Rank fusion and Top-K chunks. Insert Reranker at x=280 y=400 w=120 h=56 (shift Top-K, generator, response right by 140 and widen the viewBox to 820). Annotate the reranker's input arrow with top 150 and its output with top 20 — the winnowing ratio is the whole point.

6.9 KiB Raw Blame History Unescape Escape