pygraphistry-ai（写字术 ai）

PyGraphistry AI

Doc routing (local + canonical)

• First route with ../pygraphistry/references/pygraphistry-readthedocs-toc.md.
• Use ../pygraphistry/references/pygraphistry-readthedocs-top-level.tsv for section-level shortcuts.
• Only scan ../pygraphistry/references/pygraphistry-readthedocs-sitemap.xml when a needed page is missing.
• Use one batched discovery read before deep-page reads; avoid cat * and serial micro-reads.
• In user-facing answers, prefer canonical https://pygraphistry.readthedocs.io/en/latest/... links.

Typical workflow

1. Build graph from nodes/edges.
2. Run feature/embedding method (umap, embed, optional dbscan).
3. Inspect derived columns/features and visualize.
4. Iterate on feature columns and sampling strategy.

Baseline examples

# Similarity embedding / projection
g2 = graphistry.nodes(df, 'id').umap(X=['f1', 'f2', 'f3'])
g2.plot()

# Fit/transform flow for consistent projection on new batches
g_train = graphistry.nodes(df_train, 'id').umap(X=['f1', 'f2'])
g_batch = g_train.transform_umap(df_batch, return_graph=True)
g_batch.plot()

# Semantic search over embedded features
g2 = graphistry.nodes(df, 'id').umap(X=['text_col'])
results_df, query_vector = g2.search('suspicious login pattern')

# Text-first workflow: featurize then search/cluster
g2 = graphistry.nodes(df, 'id').featurize(kind='nodes', X=['title', 'body']).umap(kind='nodes').dbscan()
hits, qv = g2.search('credential stuffing campaign')

# Precomputed embedding columns
embedding_cols = [c for c in df.columns if c.startswith('emb_')]
g2 = graphistry.nodes(df, 'id').umap(X=embedding_cols)
g_new = g2.transform_umap(df_new, return_graph=True)

Practical guardrails

• Start with small/representative samples before full runs.
• Keep explicit feature lists (X=...) for reproducibility.
• Track engine/dataframe type for CPU vs GPU behavior.
• For anomaly workflows, document thresholds and false-positive assumptions.
• For graph ML tasks, route deeper model workflows to RGCN/link-prediction references.
• For text workflows, prefer featurize(...).umap(...).search(...) when queries are natural language.
• If users already have embeddings, reuse them via explicit embedding column lists (X=[...]) before recomputing.
• When user asks for a concise workflow snippet, prefer one short code block and avoid long narrative wrappers.

Canonical docs

• GFQL + AI combos: https://pygraphistry.readthedocs.io/en/latest/gfql/combo.html
• API AI reference: https://pygraphistry.readthedocs.io/en/latest/api/ai.html
• AI notebook index: https://pygraphistry.readthedocs.io/en/latest/notebooks/ai.html
• Example RGCN notebook: https://pygraphistry.readthedocs.io/en/latest/demos/more_examples/graphistry_features/embed/simple-ssh-logs-rgcn-anomaly-detector.html