fenic-mechanics

name

fenic-mechanics

description

Use when writing, editing, or debugging code that uses fenic (the `fenic` Python library, imported as `import fenic as fc`) — the semantic-DataFrame engine with LLM operators (semantic.map/extract/classify/predicate/reduce, embeddings), jq/markdown/transcript processing, chunking, and fuzzy matching. Covers fenic's import & namespace structure, semantic-operator calling conventions, model/session config, and known silent-failure traps. NOT for plain PySpark, pandas, or Polars code.

fenic mechanics

fenic looks like PySpark and you already write its DataFrame surface well (select/filter/join/group_by/agg, semantic.extract/classify). This skill covers the mechanics that don't transfer — where fenic differs from PySpark/pandas intuition in ways that fail (often loudly, sometimes silently). For full signatures see reference/*.md (generated from the installed version); for the correction table and traps see gotchas.md.

Golden rule: after writing or editing a fenic pipeline, run fenic check <file> — a static lint (no execution) that resolves your fc.* symbols against the installed fenic and flags namespace/import mistakes (fenic.functions, fc.array vs fc.arr, fc.explode, …). Fix what it reports.

1. Import & namespace law (the #1 source of errors)

• Always import fenic as fc. Everything hangs off fc.. There is no fenic.functions (don't write from fenic import functions as F), no fenic.api.types, and no unified OpenAIModelConfig.
• Function namespaces: fc.text.*, fc.json.*, fc.markdown.*, fc.semantic.*, fc.embedding.*, fc.dt.*, and fc.arr.* for array ops. ⚠️ fc.array(...) is a constructor for array literals; the array-operations namespace is fc.arr (fc.arr.size, fc.arr.contains, fc.arr.sort, …).
• Flat on fc: free functions (fc.col, fc.lit, fc.when, fc.coalesce, fc.count, fc.sum, fc.avg, fc.collect_list, fc.struct, fc.udf, fc.async_udf, …), all types, and all model-config classes.
• explode / unnest are DataFrame methods, not functions: df.explode("col"), df.unnest("col") — never fc.explode(...).
• PySpark camelCase aliases (withColumn, groupBy, orderBy, dropDuplicates) do exist and work, but prefer snake_case.

2. Session & models

import fenic as fc
session = fc.Session.get_or_create(fc.SessionConfig(
    app_name="my_app",
    semantic=fc.SemanticConfig(
        language_models={"mini": fc.OpenAILanguageModel(model_name="gpt-4o-mini", rpm=500, tpm=200_000)},
        default_language_model="mini",
        # embeddings are a SEPARATE class + SEPARATE dict:
        embedding_models={"emb": fc.OpenAIEmbeddingModel(model_name="text-embedding-3-small", rpm=500, tpm=200_000)},
        default_embedding_model="emb",
    ),
))

• Language vs embedding models are different classes (fc.OpenAILanguageModel vs fc.OpenAIEmbeddingModel) and live in different config keys (language_models vs embedding_models). No single unified model class.
• default_language_model / default_embedding_model are required when more than one of that kind is registered.
• Anthropic splits rate limits: fc.AnthropicLanguageModel(model_name, rpm, input_tpm, output_tpm) — no single tpm. OpenAI/Google/Cohere use tpm.
• Session creation performs a live API-key check — a valid provider key must be in the environment even just to build a semantic session.

3. Semantic operators — calling convention

• Column-level (use inside select/with_column): fc.semantic.map, extract, classify, predicate, reduce, summarize, analyze_sentiment, embed, parse_pdf.
• DataFrame-level (df.semantic.*): join (LLM predicate), sim_join (embedding similarity), with_cluster_labels (clustering). Only these three.
• Templates use Jinja2 double braces and REQUIRE matching column kwargs:

  fc.semantic.predicate("Is this a complaint? {{ msg }}", msg=fc.col("msg"))
  fc.semantic.map("Summarize {{ body }}", body=fc.col("body"))
  

  Same for fc.text.jinja(template, **columns) and df.semantic.join's predicate (which uses the literal placeholders {{ left_on }} / {{ right_on }}).
• fc.semantic.extract(col, MyPydanticModel) — schema is positional (or response_format=). fc.semantic.classify(col, [..>=2 classes..]).
• parse_pdf is fc.semantic.parse_pdf (under semantic, NOT markdown — it calls the model). Input is a column of PDF path strings (no cast needed): fc.semantic.parse_pdf(fc.col("path"), page_separator="--- PAGE {page} ---") — pass page_separator (the {page} placeholder is filled per page) when you want page breaks; omit it and pages run together.

4. ⚠️ The 4 traps fenic check can't catch

fenic check is a static lint (symbols & namespaces) — it doesn't see these. The first three run clean and produce wrong output (truly silent); the fourth errors only at execution. Get them right by hand:

1. fc.json.jq(col, query) returns an ARRAY (ArrayType(JsonType)), never a scalar. Take one match before casting: fc.json.jq(c, ".x").get_item(0).cast(fc.IntegerType).
2. Single braces in a semantic template. "... {msg} ..." (one brace) is not interpolated — the model receives the literal {msg}. Always {{ msg }}.
3. fc.dt.datediff(end, start) returns end - start. Reversed args → silently negative/wrong. Order matters.
4. fc.dt.to_timestamp / to_date / date_format take Spark/Java patterns (yyyy-MM-dd HH:mm:ss, MM-dd-yyyy), NOT Python/chrono %-tokens — fenic converts the Spark pattern to chrono internally. A %-style string raises an ExecutionError at materialization (so fenic check won't flag it). With no format, to_timestamp expects ISO-8601-with-ms; datediff/date_trunc take the resulting timestamp/date columns directly.

5. Stay in fenic — don't bypass it

Use fenic's native operators (fc.json.jq, fc.markdown.*, fc.text.parse_transcript, fc.text.extract templates, fc.text.compute_fuzzy_*) rather than dropping to json.loads, re, or manual string parsing. The point of fenic is a typed, inspectable, rerunnable pipeline — raw-Python escape hatches throw that away and don't run in the engine.

More detail

• reference/functions.md, reference/dataframe.md, reference/config-and-types.md — full signatures, generated from the installed fenic version.
• gotchas.md — the "wrote X, meant Y" correction table (every real failure mode observed) and the silent-trap deep dive.