choose-approach

Choose Approach

Decision helper. Picks the right execution lane for a text-analysis task.

When to use

• Before any skill in this plugin is applied to more than a few hundred documents.
• When the user says "analyze all my notes", "categorize everything", "run NER on the corpus" — i.e. any whole-corpus pass.
• When cost could exceed a few dollars if done naively.

The three lanes

Lane	Strengths	Weaknesses	Cost model
Classical NLP (spaCy, scikit-learn, gensim, TextBlob, BlackLab)	Deterministic, fast, free, scales to millions of docs, strong for frequency/NER/TF-IDF/LDA	Brittle on short or noisy text; no semantic judgment	CPU time only
Local LLM (Ollama: llama3.1, qwen2.5, gemma2)	No per-token cost, private, good for classification/labeling on small batches	Slow on big corpora (GPU-bound), lower ceiling than frontier models	Electricity + wall-clock time
Cloud LLM (OpenRouter → Claude, GPT, Gemini, DeepSeek, Llama)	Best judgment, handles nuance, parallelizable	Per-token cost — dangerous on 10k+ docs without planning	$ per M tokens

Procedure

1. Size the corpus: count documents, estimate avg tokens/doc (approx words × 1.3), compute total input tokens.
2. Classify the task as one of:
   • Mechanical (frequency, length, NER with known list, metadata corr) → Classical NLP
   • Semantic judgment (categorize, taxonomy, synonym-cluster, topic-label) → LLM (local or cloud)
   • Hybrid (candidate extraction + labeling) → Classical NLP first, LLM on reduced set
3. If LLM: estimate cost at three tiers using current OpenRouter pricing (always fetch live — prices change):
   • Cheap: DeepSeek V3 / Gemini Flash / Llama 3.3 70B (~$0.05-0.30 per M input tokens)
   • Mid: Claude Haiku, GPT-4o-mini (~$0.25-1.00 per M)
   • Premium: Claude Sonnet/Opus, GPT-4o (~$3-15 per M)
4. Apply cost-reduction tactics before running:
   • Sample, don't sweep — 500 stratified docs usually derive the same taxonomy as 50,000.
   • Two-pass: cheap model labels everything, premium model re-labels only the low-confidence residue.
   • Map-reduce: summarize per-doc cheaply, then cluster summaries.
   • Embeddings + clustering instead of LLM for grouping: one embedding call per doc (very cheap) + k-means is orders of magnitude cheaper than asking an LLM to compare docs.
   • Prompt caching (Anthropic, OpenAI): if the same system prompt + taxonomy ships with every call, cache it.
5. Present the user with: recommended lane, estimated cost range, and at least one cheaper fallback. Ask for go-ahead before any run >$2.

Output shape

Task: <restatement>
Corpus: N docs, ~M tokens total
Recommended lane: <Classical NLP | Local LLM | Cloud LLM via OpenRouter>
Recommended model (if LLM): <name> — est. $X.XX for full corpus
Cheaper fallback: <sampling strategy OR cheaper model> — est. $Y.YY
Reasoning: <2-3 sentences>

Hard rules

• Never launch an LLM pass over >1k documents without showing a cost estimate first.
• Never use a premium model for a task a cheap model can do (labeling into known categories, routing).
• For pure mechanical tasks, never suggest an LLM. spaCy/regex is the right answer.