Overview

Sixteen commits, three threads: a new HEX-only injection mode (the tone-image is dropped, only the hex palette is injected into the prompt), an angle picker split into three categories with an inline UI, and a Korean-prompt attribute extractor that stops the prod regression where “하늘을 달리는 남자” got a female model attached. A small OTLP tuning pass (batched spans, widened metric interval) wraps it up.

Previous post: hybrid-image-search-demo Dev Log #16

graph LR
    U["User Prompt (KO)"] --> Extract["Attribute Extractor
spaCy + few-shot LLM"]
    Extract --> Gender["gender"]
    Extract --> Age["age"]
    Extract --> Race["race"]
    Gender --> Model["Model image injection"]
    Age --> Model
    Race --> Model
    U --> Inject{Injection Mode}
    Inject -->|off| P0["Plain prompt"]
    Inject -->|auto| PT["Tone image + hex"]
    Inject -->|hex_only| PH["Hex palette only"]
    Model --> Prompt["Final prompt"]
    P0 --> Prompt
    PT --> Prompt
    PH --> Prompt

HEX-Only Injection Mode

The tone-injection system has two axes: the tone reference image (3- or 5-image pack, extracted to hex colors) and the prompt fragment that frames the image inside the generation prompt. The existing default wired both — inject the image and the tone-direction text. The ask in this session was a third mode where only the hex palette goes into the prompt as color guidance, and the tone image path is skipped entirely.

The design (44d5bff, 08916cb) unified this as a three-way injection_mode enum: off / auto / hex_only. Plumbing it through was the bulk of the work:

• refactor(prompt): promote hex_colors to an explicit param and add a hex-only prompt block (0a16f4f).
• feat(db): thread injection_mode through log_generation and hydration (e53be41) — necessary so the generation record can reconstruct the mode later for debugging.
• feat(backend): wire injection_mode end-to-end off/auto/hex_only (e6807e2).
• Alembic migration 20260420_add_injection_mode.py adds the column.
• feat(ui): three-way toggle pill, with a11y tooltips explaining each mode (5659fd3, 3b2cf22).

The UI polish (51464e6) took a few iterations. The initial design was neutral-gray for inactive states, but “off” didn’t read as disabled — users kept thinking it was still active. The fix: inactive pills turn red, the active pill yellow. Stronger contrast, the state legible at a glance.

A fix(ui) commit (988ea37) covers the prompt-display helpers that were still showing the tone-direction text even in hex_only mode — a dangling copy path. And chore(gen) (c419349) polished the telemetry labels so the three modes show up clearly in Grafana spans.

Angle Picker: Three Categories with Inline UI

Commit 61c5802 splits angle selection from a flat list into three categories with an inline UI (presumably “general / beauty / product” or similar, based on the prior Lens picker pattern in #16). The structural motivation is the same as the lens picker expansion two logs back: flat lists over ~5 items become noise; grouping restores scanability.

The subtle frontend concern was that the three-category split needed a deterministic display order — reflectiveness in the backend’s angle_registry plus category metadata in the JSON schema. The component reads the schema once and renders sections; selection still emits a single angle_id to the backend, so the API surface is unchanged.

Korean Prompt Attribute Extraction

The prod bug that kicked this off: a prompt “하늘을 달리는 남자” (a man running across the sky) produced a generation where the model-injected reference image was Araya 05.png, which is labeled female in data/model_labels.json. The LLM-driven attribute extractor was picking the wrong gender.

The fix (61e6c85) is a few-shot prompt that enforces gender/age/race extraction with examples. Tightening the prompt schema rather than running a classifier keeps the solution simple — the decision made in-session was that a minor guardrail was enough, since the input space of Korean prompts is broad and any real classifier would need a labeled corpus.

spaCy pinning (9f2773b) is related — en_core_web_sm was auto-upgrading in fresh venvs, and the prompt parser relies on specific token types. Pinning ensures reproducible parses.

OTLP Telemetry Tuning

Two small but load-bearing changes (02c0c6c): batch spans instead of sending per-span (one of the OpenTelemetry defaults that absolutely needs overriding under any real traffic), and widen the metric interval so Grafana Cloud free-plan ingestion stays well under limits. The trial expired — the dashboards need to fit in free.

Commit Log

Insights

Two threads share the same lesson: explicit modes beat implicit fallbacks. The injection_mode enum is strictly better than the prior “the flag defaults affect two things simultaneously” design, because each code path is now legible at the call site — no need to trace through eight booleans. Similarly, the Korean prompt extractor used to rely on an LLM’s default behavior, which worked most of the time until it didn’t; a few-shot prompt is still LLM-based, but now the decision is visible in the prompt itself. Visual contrast follows the same principle: the moment an “off” toggle looks neutral, users stop reading it as off. Next session’s likely focus: the auto-fill token expansion request from session 4, which needs a thoughtful UI for editing 3 or 5 tone refs at once instead of one-at-a-time.