Llm feature engineering

LLM Feature Engineering

An LLM is a non-deterministic dependency — it can be wrong, slow, expensive, or manipulated. Treating it like fetchUser(id) produces flaky, unsafe features. Engineer around its nature: define a contract, constrain output, ground facts, measure quality with evals, handle failure modes, and never trust raw model text as correct ([[hardening]]).

LLM features are products — users need predictable behavior, fallbacks when AI fails, and clarity when output is machine-generated. "Demo magic" is not shipping.

Pairs with [[context-curation]] for prompt payload size, [[source-first]] for grounding, [[interface-design]] for tool schemas, [[resilience]] for timeouts/retries, [[observability]] for cost/quality/latency, [[test-first]] mindset for eval sets, [[incremental-delivery]] to ship slices, and [[launch-readiness]] for gradual rollout of AI behavior.

When to Use

• Adding AI: generation, summarization, extraction, classification, chat, agents, copilots
• Designing prompts, system instructions, RAG pipelines, or tool/function definitions
• A model-powered feature is inconsistent, hallucinating, slow, or too expensive
• Putting an LLM feature into production or preventing quality regression after prompt/model changes
• Reviewing AI-related PRs for safety and contract discipline

Also ask: should this be an LLM at all? Rules, regex, search, or classical ML often beat a general model for narrow tasks — cheaper, deterministic, auditable. Use LLM when language understanding, open-ended generation, or flexible reasoning is genuinely required.

Skip heavy LLM engineering for one-off internal scripts with no user impact — still validate outputs if they touch production data.

Process

Work in order. Prompt tuning before task definition is wasted iteration.

1. Define the task contract — before prompts

Write what the feature must do in testable terms ([[spec-first]]):

Field	Example
Input	User question + retrieved docs; ticket text; form fields
Output	JSON { summary, category, confidence }; user-visible reply; tool call only
Success	≥90% category match on eval set; summary contains all key facts; p95 < 3s
Failure cost	Wrong medical dose = critical; wrong emoji suggestion = low
Out of scope	Model must not execute purchases, access arbitrary data

Vague goals ("be helpful") are unmeasurable. One sentence output contract minimum.

Choose pattern:

Pattern	Use when	Avoid when
Classification / extraction	Fixed labels, fields from text	Need creative prose
Generation	Drafts, explanations, transforms	Facts without sources
RAG	Answers must cite company data	Tiny static FAQ (use search)
Tool/agent	Multi-step actions in product	Simple single API call
Hybrid	Retrieve → extract → generate with schema	Over-agent for one lookup

2. Pick model and budget — fit task to model

Don't default to the largest model:

Factor	Tradeoff
Quality	Hard reasoning, long context → larger model
Latency	Chat UX → smaller/faster or streaming
Cost	High volume → smaller model + better prompt/RAG
Determinism	Lower temperature for extraction; higher for creative drafts

Set token budgets — max input/output tokens, max retrieval chunks ([[context-curation]]). Instrument cost per request from day one ([[observability]]).

Re-evaluate when volume grows — a cheap model at scale beats a premium model on every request.

3. Structure the prompt — role, task, format, constraints

System prompt — stable rules: role, safety boundaries, output format, what never to do.

User/content — task instance; separate from untrusted data when possible ([[hardening]]).

Effective structure:

Role: You extract order issues from support tickets.

Task: Return category and one-line summary.

Output: JSON only, schema: { "category": enum, "summary": string, "confidence": 0-1 }

Rules:
- If unsure, category "unknown" and confidence < 0.5
- Do not invent order IDs not in the ticket
- Max summary 200 chars

Examples:
  Input: "..." → Output: { ... }

Few-shot examples — 2–5 representative cases, including edge cases (empty input, ambiguous). Update examples when evals fail — don't grow prompts without measurement.

Keep prompts minimal ([[context-curation]]) — every token is cost, latency, and distraction.

4. Constrain and validate output — never trust raw text

Model text is untrusted input to your app ([[hardening]]):

• Structured output — JSON Schema, tool calls, enums — request and parse strictly
• Validate after generation — schema validator, allowlist categories, regex for IDs
• Repair loop — one retry with "your JSON was invalid: …" — cap retries ([[resilience]])
• Never eval(), JSON.parse without schema, or SQL from model output without guardrails
• Never execute shell/code from model without sandbox and allowlist

const parsed = schema.safeParse(modelOutput);
if (!parsed.success) return fallbackOrRetry(parsed.error);
// use parsed.data only

Free-form text to users is OK for display — still sanitize HTML if rendered; don't pipe to downstream systems without validation.

5. Ground factual tasks — RAG and citations

For facts about your product, docs, or user data — model memory hallucinates ([[source-first]]):

1. Retrieve relevant chunks (search, embeddings) — don't stuff whole knowledge base
2. Pass sources in prompt with clear delimiters
3. Require citations — chunk id, URL, or quote — validate cited ids exist
4. Answer only from sources instruction — with fallback "I don't have that information"

RAG quality = retrieval quality. Eval retrieval separately from generation (did we fetch the right doc?).

Refresh index when source data changes; stale RAG is silent wrong answers.

6. Tools and agents — minimal power, full authz

Tools turn models into action surfaces — highest risk ([[interface-design]], [[hardening]]):

• Allowlist tools — model cannot call APIs you didn't define
• Strict JSON schemas for arguments — validate server-side
• Same authorization as if the user clicked the button — tenant, role, object ownership
• Idempotent writes where retries happen ([[resilience]])
• Confirm destructive actions in product UI — don't let model alone delete data
• Cap tool loops — max steps per request; timeout entire agent run

Bad tool: run_sql(query: string)
Good tool: search_orders(userId, status, limit) — scoped, typed

Log tool calls with correlation id for audit ([[observability]], [[decision-docs]] for sensitive flows).

7. Build an eval set — your non-deterministic test suite

Without evals, prompt changes are guesswork ([[test-first]] mindset):

Golden set — 20–100+ real or realistic inputs with expected or rubric-scored outputs:

Task type	Eval approach
Classification	Exact match or confusion matrix
Extraction	Field-level match; fuzzy for text
Generation	Rubric (human or LLM-judge), key facts present
RAG	Answer correct + citation correct

Run evals:

• Before/after every prompt or model change — no ship without comparison
• In CI for schema/regression checks; full quality eval nightly or pre-release
• Track metrics over time — quality drift as models update upstream

Store eval cases in repo — version with prompts. Failed cases become regression additions.

8. Handle failure modes explicitly

Decide behavior for each — no silent undefined:

Failure	Product behavior
Timeout	Bounded wait ([[resilience]]); fallback message or cached response
Rate limit / 5xx	Retry with backoff; degrade gracefully
Malformed output	Retry once → fallback
Refusal	Explain limitation; offer non-AI path
Low confidence	Don't auto-apply; human review queue
Model unavailable	Feature off or rule-based backup
Prompt injection	Don't follow user instructions to exfiltrate; layer defenses ([[hardening]])

Fallbacks should be designed, not "show error." User still completes the task.

Surface failures to telemetry — silent fallback hides systemic quality issues.

9. Defend against prompt injection

Any untrusted text in the prompt (user message, webpage, email, ticket) can try to override instructions:

Layers (none alone is sufficient):

• Separate system rules from user content structurally (API message roles)
• Delimiters — ### User input ### — weak alone but helps
• Output contract — model must return schema; ignore "ignore previous instructions" in prose output
• Tool allowlist — can't exfiltrate via tools not exposed
• Don't put secrets in prompts — model may leak in output
• Monitor for anomalous tool calls or data exfil patterns

For high-risk (PII, payments, admin): human approval on actions; model proposes only.

10. Instrument production — quality, cost, latency

Ship with dashboards ([[observability]]):

• Latency p50/p95 per stage — retrieve, model, tools
• Token usage — input/output/cached; cost per feature
• Quality proxies — thumbs down, edit rate, human override rate, task completion
• Error rates — schema validation failures, timeouts, refusals
• Safety — blocked injections, policy violations

Alert on user-visible degradation — error spike, latency SLO, quality metric drop — not every failed parse.

Sample production traces for periodic human review — eval set drifts from reality.

11. Ship and iterate safely

Incremental delivery ([[incremental-delivery]], [[launch-readiness]]):

1. Internal dogfood with logging
2. Shadow mode — run model, don't show output; compare to baseline
3. Beta cohort or feature flag
4. Gradual ramp with quality/cost watch
5. Prompt/model changes gated on eval improvement

Human-in-the-loop where stakes are high — support suggestions, not auto-sent replies; clinician review for health; editor for published content.

Document AI disclosure where product/policy requires — users know output is machine-generated.

Version prompts and models in config — know what's live when debugging regressions.

12. Scenario playbooks

Support ticket classification

Schema output → eval on labeled tickets → confidence threshold → route low confidence to human.

Doc Q&A (RAG)

Index pipeline → retrieval eval → generation with citations → validate cited chunks → fallback "no answer".

Chat copilot

Stream UX → token budget → injection layers → no tool without authz → rate limit per user.

Extraction to downstream API

Extract JSON → validate → map to API contract ([[interface-design]]) → never skip validation.

Agent with 3 tools

Allowlist → max 5 steps → log each call → timeout 30s → fallback to manual flow.

Summarization

Length limits → don't summarize secrets into output → eval key facts preserved.

Model upgrade

Same eval set on new model → compare cost/latency/quality → staged rollout.

Prompt hotfix in prod

Eval diff required → flag-gated deploy → watch override rate 24h.

Common Rationalizations

• "The prompt works on my examples." — Two examples isn't an eval; the third run fails differently.
• "The model is smart, it'll figure it out." — Underspecified tasks drift and hallucinate.
• "We'll parse the text." — Unvalidated free text breaks downstream; constrain and validate.
• "Prompt injection won't happen to us." — Untrusted text in the prompt is an attack surface.
• "Use the biggest model to be safe." — Often slower, pricier, with no quality gain on narrow tasks.
• "We'll add evals after launch." — You can't regression-test what you never measured.
• "Temperature 0 means deterministic." — Still non-deterministic; still need evals and validation.
• "RAG means always accurate." — Bad retrieval → confident wrong answers.
• "Agents are the future — use agents." — Agents multiply cost, latency, and failure modes.
• "Users love AI." — Users love working features; wrong AI erodes trust fast.

Red Flags

• No eval set — quality judged by eyeballing a couple of runs
• Raw model output executed, parsed without schema, or passed to SQL/shell
• User/web/email content concatenated into prompt with no separation or defenses
• Tools with arbitrary SQL/shell/file access
• No fallback when model errors, refuses, or returns garbage
• Unbounded context, retrieval, or agent steps — cost and latency explosions
• Prompt changes shipped without before/after eval metrics
• No production metrics on quality, cost, or latency
• High-stakes action (payment, delete, send email) without human confirmation
• Citations not validated — model cites missing sources
• Same prompt in prod with no version id — can't debug regressions
• "AI" used where rules or search would be deterministic and cheaper

Verification

• Task contract defined — input, output shape, success metrics, failure cost ([[spec-first]])
• LLM chosen deliberately — not default largest model; token budgets set
• Prompt structured — role, format, constraints, few-shot examples; context minimal ([[context-curation]])
• Output structured and validated server-side; no blind trust of text ([[hardening]])
• Factual tasks grounded — RAG/retrieval with citation checks ([[source-first]])
• Tools allowlisted, schemas strict, authz matches HTTP API ([[interface-design]], [[hardening]])
• Eval set exists; prompt/model changes gated on measured improvement ([[test-first]] mindset)
• Failure modes handled — timeout, retry, malformed, refusal, unavailable ([[resilience]])
• Injection defenses layered; high-stakes actions require human approval
• Cost, latency, quality instrumented in production ([[observability]])
• Gradual rollout plan — flag, beta, monitoring ([[launch-readiness]])
• Prompt/model versioned; disclosure and fallbacks documented for users