LLM Analysis

Classify, summarize, enrich, and deduplicate data using AI models

Overview

Anysite CLI integrates with LLM providers to add AI-powered analysis to dataset pipeline data. Six operations are available: classify, summarize, enrich, generate, match, and deduplicate. Every operation reads records from a dataset source's collected Parquet files — there is no --file input for arbitrary JSON.

Requires the llm extra: pip install "anysite-cli[llm]"

Setup

# Interactive
anysite llm setup

# Non-interactive (CI/agents)
anysite llm setup --provider openai --api-key sk-... --no-test
anysite llm setup --provider anthropic --api-key-env ANTHROPIC_API_KEY

--api-key accepts the key value directly; --api-key-env references an environment variable instead. --model overrides the default model; --no-test skips the connection test.

Supported Providers

Provider	Default Model
OpenAI	`gpt-4.1-mini`
Anthropic	`claude-sonnet-4-5-20250514`

Provider settings are stored under llm.providers.<name> in ~/.anysite/config.yaml.

Operations

Every operation shares this set of options: --provider, --model, --parallel/-j (default 5), --rate-limit, --temperature, --max-tokens, --no-cache, --format/-f, --output/ -o, --quiet/-q, --prompt, --prompt-file, and --dry-run (print the built prompt without calling the LLM).

Classify

anysite llm classify dataset.yaml --source profiles \
  --categories "developer,recruiter,executive,other" \
  --fields "name,headline,summary"

If --categories/-c is omitted, the LLM auto-detects 3-7 categories from the first 20 records. --multi allows more than one category per record. --output-column (default category) names the result column.

Summarize

anysite llm summarize dataset.yaml --source profiles \
  --fields "name,headline,summary,experience" \
  --max-length 50 \
  --output-column bio_summary

--max-length (default 100) caps summary length in words.

Enrich

anysite llm enrich dataset.yaml --source profiles \
  --add "seniority:junior/mid/senior/lead" \
  --add "is_technical:boolean" \
  --add "years_experience:number" \
  --add "primary_skill:string"

Supported attribute types in --add name:spec:

Enum — slash-separated choices: "seniority:junior/mid/senior"
Boolean — "is_technical:boolean"
Number — "years_experience:number"
Integer range — a spec containing a hyphen, e.g. "quality_score:1-10", is treated as an integer
String — free text: "primary_skill:string"

Generate

anysite llm generate dataset.yaml --source profiles \
  --prompt "Write a 2-sentence professional intro for {name} who works as {headline}" \
  --temperature 0.7 \
  --output-column intro_text

Either --prompt or --prompt-file is required. --output-column defaults to text. If --temperature is not given, generate defaults to 0.7; every other operation defaults to 0.0.

Match

anysite llm match dataset.yaml \
  --source-a profiles \
  --source-b companies \
  --top-k 3

Field selection for match uses --fields-a and --fields-b (not the shared --fields). --top-k (default 3) caps matches per source-a record; --threshold (default 0.5) is the minimum match score to keep.

Deduplicate

anysite llm deduplicate dataset.yaml --source profiles \
  --key name \
  --threshold 0.8

--key/-k (default name) blocks records before comparison — only records sharing the first three characters of this field (lowercased) are compared, to limit LLM calls on large sources. --threshold (default 0.8) is the minimum confidence to flag a pair as duplicates.

Using LLM in Dataset Pipelines

sources:
  - id: profiles
    endpoint: /api/linkedin/user
    from_file: users.txt
    input_key: user

  - id: profiles_enriched
    type: llm
    dependency:
      from_source: profiles
      field: name
    llm:
      - type: classify
        categories: "developer,recruiter,executive,sales,other"
        output_column: role_type

      - type: enrich
        add:
          - "seniority:junior/mid/senior/lead"
          - "is_technical:boolean"

      - type: summarize
        max_length: 50
        output_column: bio_summary

Multiple LLM steps can be chained within a single LLM source. They execute in order, each adding new columns to the dataset.

Caching

Responses are cached in ~/.anysite/llm_cache.db to avoid repeated API calls:

anysite llm cache-stats
anysite llm cache-clear
anysite llm classify dataset.yaml --source profiles \
  --categories "dev,recruiter,exec" --no-cache

Options Reference

Option	Description	Applies To
`--source`, `-s`	Source to process	all (except match, which uses `--source-a`/`--source-b`)
`--fields`	Fields to include in the LLM prompt (comma-separated)	classify, summarize, enrich, deduplicate, generate
`--fields-a` / `--fields-b`	Fields from each source	match
`--categories`, `-c`	Comma-separated categories	classify
`--multi`	Allow multiple categories per record	classify
`--add`	Attribute spec `name:type_or_values` (repeatable)	enrich
`--prompt` / `--prompt-file`	Custom or file-based prompt template	all
`--temperature`	LLM temperature (0.0-1.0)	all
`--max-tokens`	Max response tokens	all
`--max-length`	Max words for the summary	summarize
`--output-column`	Name for the result column	classify, summarize, generate
`--top-k`	Number of matches per record	match
`--key`, `-k`	Field to block/compare on	deduplicate
`--threshold`	Match or duplicate confidence threshold (0.0-1.0)	match, deduplicate
`--provider` / `--model`	Override the configured LLM provider or model	all
`--dry-run`	Print the built prompt without calling the LLM	all
`--no-cache`	Skip the LLM cache	all

Next Steps

SQL Querying

Query and analyze your enriched data with DuckDB SQL

Examples

See complete end-to-end workflow examples