Use Case: LLM Cost Optimization

What you’re solving for

Your team or product spends measurable money on LLM API calls for code questions. You need to reduce spend without moving to a cheaper model or cutting features, and report the savings to a stakeholder.

Step 1 — Baseline the current spend

Before installing NeuralMind, capture what you’re spending. Pick a representative workday:

# Count tokens per query today by logging your agent's input/output
# (most agents have a debug mode or you can estimate with tiktoken)

Compute: avg_tokens_per_query × queries_per_day × 30 × $_per_MTok. That’s your monthly floor.

Step 2 — Install NeuralMind

pip install neuralmind
neuralmind build .
neuralmind install-hooks .     # Claude Code users only

Step 3 — Measure the new baseline

neuralmind benchmark . --json

Returns:

{
  "wakeup_tokens": 341,
  "avg_query_tokens": 739,
  "avg_reduction_ratio": 65.6,
  "results": [...]
}

Compare avg_query_tokens to your pre-install baseline. This is the retrieval-side savings.

Step 4 — Measure consumption-side savings (Claude Code)

PostToolUse hooks compress Read/Bash/Grep output. Rough numbers:

Tool	Typical reduction
Read	~88% (file → skeleton)
Bash	~91% (errors + tail)
Grep	Capped at 25 matches

Combined retrieval + consumption is typically 5–10× total reduction vs baseline.

Step 5 — Report to stakeholders

v0.45.0+: let NeuralMind write the dollar figures for you. The manual tokens × price arithmetic below is now built in — neuralmind savings --cost reads your opt-in event log and prices the measured savings on input tokens:

neuralmind savings . --cost --model claude-opus-4-8 --queries-per-day 100
# →   Dollar savings — claude-opus-4-8 @ $5.0/MTok input
# →     Cost without NM : $    829.25
# →     Cost with NM    : $      3.47
# →     Saved           : $    825.78
# →     Projected       : $24.90/day · $746.86/month  (at 100 queries/day)

--model picks from a built-in input-price table (Claude / GPT / Gemini, snapshot 2026-07), --queries-per-day sets the projection assumption, and --cost --json emits a dollar_savings block you can paste into a dashboard or the template below. These are measured-usage estimates (anchored to the 50k-tokens/query baseline the token report already discloses), not an invoice reconciliation — for the stakeholder one-pager that’s exactly the framing you want, with every assumption printed next to its number.

A one-page summary template:

NeuralMind rollout — token cost impact

Baseline: {avg_tokens} × {queries/day} × 30 × ${price}/MTok = ${monthly}

After NeuralMind: {new_tokens} × {queries/day} × 30 × ${price}/MTok = ${new_monthly}

Reduction: {ratio}× on retrieval, {total_ratio}× combined with PostToolUse hooks

Setup cost: one-time neuralmind build (~minutes)

Ongoing cost: incremental rebuild on git commit (seconds)

Risk: fully local, no new SaaS dependency, MIT-licensed

Ongoing hygiene

Index freshness: neuralmind init-hook . auto-rebuilds on every commit.
Adapt to workflow: enable memory (TTY prompt) and install the hooks (neuralmind install-hooks .) so the synapse layer learns from your actual usage automatically — no manual step, and stale associations decay instead of lingering.
Model changes: run neuralmind benchmark again when you switch models — absolute dollar savings scale with input price.

Debugging cost spikes with the graph view (v0.6.0+)

If a query returns 5K tokens when you’d expect 800, you used to be debugging by reading log files. v0.6.0 makes it visual.

neuralmind serve . in a separate terminal.
Run the offending query.
In the detail panel, click Replay last query.

The graph view highlights the L3 hits the agent received. The diagnosis is usually obvious from the pulse pattern — a stale cluster boundary that grabbed too many nodes, a missing structural edge that forced the retriever wider, or an unexpected hub node pulling in unrelated context. Fix the underlying issue (rebuild the index, update CLAUDE.md, or tune the cluster boundary) and the next replay shows a tighter result.

The pulse-rings live feed is also useful during normal use: if you notice the canvas going quiet during sessions you’d expect to be busy, that’s a signal the agent isn’t actually using NeuralMind retrieval (maybe the MCP server isn’t wired up, or the hooks didn’t install). A visual heartbeat is faster than checking a log.

What this doesn’t fix

Output tokens — NeuralMind reduces input context, not model output length. Pair with prompt instructions to keep responses concise.
Non-code LLM usage — docs QA, ticket triage, etc. NeuralMind is code-specific.
One-off experiments — savings compound with repeated queries; single questions show less benefit.

← Back to use-case index · Main README