topoteretes--cognee/cognee/eval_framework/token_usage_analysis

token_usage_analysis

Estimate the token cost of cognee persistent memory versus full-context prompting, from a few representative chunks of any text.

This is a follow-up to the earlier token-cost probe, which measured that cost on a single synthetic corpus. Here we characterize how it varies with the input: running the same measurement across a spectrum of real text — fiction, news, encyclopedic — and across chunk sizes, so the break-even can be read for text that resembles a given workload rather than one fixed corpus.

The measurement centers on ingestion cost, which is dominated by graph extraction. Its output-tokens-per-input-token ratio scales with how densely entities and relations are packed into the text, and that ratio in turn drives the break-even. Running it over a density spectrum shows the two moving together — a small, concrete extension of the note's cost model to a wider range of inputs.

The cost model

Two strategies answer the same repeated queries over a fixed corpus. Each one's cumulative token cost is a straight line in the number of queries:

full-context:  queries × (corpus_tokens + query_overhead)
cognee:        ingestion_tokens + queries × retrieved_context

• ingestion_tokens is the one-time cost of cognee.remember() — summarizing and graph-extracting every chunk. We measure it on a few representative chunks and scale by the corpus size: ingestion_tokens = multiplier × corpus_tokens, where multiplier is measured ingestion tokens per content token.
• retrieved_context is the (roughly constant) context cognee.recall() packs per query.

Reading off this model, each reduction milestone answers: after how many repeated queries does the cumulative full-context cost become factor× the cumulative cognee cost, including cognee's one-time ingestion cost?

queries(factor) =
    factor × ingestion_tokens
    / ((corpus_tokens + query_overhead) − factor × retrieved_context)

Parity, the cross-over where both cumulative costs are equal, is factor = 1. For example, the 7 milestone is the query count where full-context has spent 7× as many input/context tokens as cognee, so cognee uses 7× fewer input tokens at that point. No milestone is privileged — break-even is just one of them.

Token usage is read from the real LLM responses (prompt + completion), so the instruction wrapper and the Pydantic graph schema are included with no estimated constants.

Results: the density spectrum

Measured with openai/gpt-5-mini, 3 sampled chunks per corpus, retrieved_context = 1118, query_overhead = 32. Each corpus is ~10–12k tokens. We ran the whole spectrum at two chunk sizes to show that chunk size, not just density, moves the numbers: cognee's 8191 default and 4095 (half). Values are graph output ÷ input / ingestion multiplier / parity (queries).

Tier	Source	Corpus tokens	at chunk 8191	at chunk 4095
Fiction	War and Peace excerpt	10,504	2.8× / 5.7× / ~6	4.5× / 8.0× / ~9
News	Wikinews (23 articles)	10,161	4.2× / 7.3× / ~8	5.9× / 9.8× / ~11
Encyclopedic	Wikipedia — Apollo 11	10,037	3.4× / 6.5× / ~7	5.9× / 10.3× / ~12
Dense synthetic	benchmark people_raw	12,009	9.3× / 12.0× / ~13	10.1× / 14.0× / ~15

Three things stand out:

• The ingestion multiplier tracks the graph-output ratio, and the break-even in queries tracks the multiplier — exactly as the cost model predicts. The dense synthetic corpus produces ~9–10× graph output per input token; ordinary prose produces ~3–6×.
• Realistic text is clustered and lower. Fiction, news, and encyclopedic prose break even in the single digits to low teens; the dense synthetic corpus is the outlier. The technical note's ~23–26 queries reflects that dense outlier (and larger ingestion chunks), not typical input.
• Chunk size matters too. Halving the chunk size raises every ratio and multiplier, because smaller chunks carry proportionally more fixed prompt/schema overhead per content token. Same corpus, same model — different break-even.

Caveat: these are single runs, and graph extraction is not fully deterministic — its output length varies between runs (most on the dense corpus, where a repeat of the 8191 run moved the multiplier by several points). Treat the numbers as representative, not exact; the ordering is the robust result.

Per-tier cumulative-cost plots and full per-chunk JSON are in results/chunk_8191/ and results/chunk_4095/.

Usage

Install the project with the eval dependencies, then run the commands from this directory:

uv sync --dev --all-extras
cd cognee/eval_framework/token_usage_analysis

The script loads the repo-root .env, so make sure it contains a working LLM_PROVIDER, LLM_MODEL, and API key. If --llm-models is omitted, the configured LLM_MODEL is used.

# one representative file (the input is treated as the corpus)
uv run python analyze.py --file data/wikipedia_article.txt --plot

# a folder of .txt files, pooled then sampled
uv run python analyze.py --dir some_corpus/ --out report.json

# a single representative chunk (corpus size must be given explicitly)
uv run python analyze.py --text "$(cat one_chunk.txt)" --corpus-tokens 200000

The script chunks the input with cognee's TextChunker, samples a few chunks (--samples, default 3), measures each through cognee.remember()'s summary and graph-extraction calls, and writes a JSON report. --plot additionally writes the cumulative-cost figure (matplotlib is included in the evals extra).

Key options:

Flag	Default	Meaning
--file / --dir / --text	—	input form (exactly one)
--samples	3	chunks to measure
--llm-models	the .env model	comma list; runs each, switching cognee's config
--corpus-tokens	token count of the input	corpus size for the comparison (required with --text)
--retrieved-context	1118	recall context per query
--query-overhead	32	instruction + question tokens per query
--reduction-factors	1,2,7,10	milestones to report (1 = parity)
--plot / --plot-dir	off / .	also write the cross-over figure

Reproducing the spectrum

The spectrum is four single-input runs per chunk size, collated by hand. Each chunk size gets its own results folder:

for size in 8191 4095; do
  for tier in war_and_peace wikinews wikipedia dense; do
    uv run python analyze.py --file "data/${tier}"*.txt --max-chunk-size "$size" \
      --out "results/chunk_${size}/${tier}.json" --plot --plot-dir "results/chunk_${size}"
  done
done

--max-chunk-size defaults to 4095; pass 8191 for cognee's default.

--llm-models runs only models you have keys for. To compare providers, add OPENAI_API_KEY / ANTHROPIC_API_KEY to .env; the script switches LLM_PROVIDER, LLM_MODEL, and LLM_API_KEY before each measurement. The runs above used the configured openai/gpt-5-mini.

Modules

analyze.py is pure orchestration; each step is one call into a focused module:

• cli.py — argument definitions and default resolution.
• corpus.py — input → sampled chunks (cognee TextChunker). No LLM, no math.
• measure.py — the only LLM/IO surface: run a chunk through cognee and record real token usage per model.
• cost_model.py — pure arithmetic: the two strategy cost objects, the average chunk, ingestion scaling, and reduction milestones.
• report.py — per-model orchestration and JSON assembly.
• plot.py — optional cumulative-cost figure (matplotlib, imported lazily).

Assumptions and caveats

• Constant retrieved context. Recall context is treated as fixed per query (it grows only modestly with corpus size under a fixed top_k).
• Embeddings excluded. Only language-model tokens are counted; embeddings are computed locally.
• Representative-chunk extrapolation. Ingestion is scaled from the average of the sampled chunks, assuming the rest of the corpus has similar density.

Data and attribution

The data/ excerpts are the raw text that actually gets ingested:

• war_and_peace_excerpt.txt — War and Peace, Project Gutenberg eBook #2600 (public domain).
• wikipedia_article.txt — Apollo 11, English Wikipedia (CC BY-SA 4.0). https://en.wikipedia.org/wiki/Apollo_11
• wikinews_article.txt — 23 recent English Wikinews articles, concatenated to a comparable corpus size (CC BY 2.5). Titles and URLs in data/wikinews_sources.json.
• dense_synthetic.txt — an excerpt of the benchmark people_raw corpus shipped in this repo.