flockmtl

Downloads

499

GitHub-Sterne

267

Erweiterungs-Repository auf GitHub

LLM- und RAG-Erweiterung zur Kombination von Analytik und semantischer Analyse

Installation und Laden

INSTALL flockmtl FROM community;
LOAD flockmtl;

Beispiel

-- After loading, any function call will throw an error if the provider's secret doesn't exist

-- Create your provider secret by following the [documentation](https://dais-polymtl.github.io/flockmtl/docs/what-is-flockmtl/). For example, you can create a default OpenAI API key as follows:
D CREATE SECRET (TYPE OPENAI, API_KEY 'your-api-key');

-- Call an OpenAI model with a predefined prompt ('Tell me hello world') and default model ('gpt-4o-mini')
D SELECT llm_complete({'model_name': 'default'}, {'prompt_name': 'hello-world'});
┌──────────────────────────────────────────┐
│ llm_complete(hello_world, default_model) │
│                 varchar                  │
├──────────────────────────────────────────┤
│                Hello world               │
└──────────────────────────────────────────┘

-- Check the prompts and supported models
D GET PROMPTS;
D GET MODELS;

-- Create a new prompt for summarizing text
D CREATE PROMPT('summarize', 'summarize the text into 1 word: {{text}}');

-- Create a variable name for the model to do the summarizing
D CREATE MODEL('summarizer-model', 'gpt-4o', 'openai');

-- Summarize text and pass it as parameter
D SELECT llm_complete({'model_name': 'summarizer-model'}, {'prompt_name': 'summarize','context_columns': [{'data': 'We support more functions and approaches to combine relational analytics and semantic analysis. Check our repo for documentation and examples.'}}]);

Über flockmtl

FlockMTL ist eine experimentelle DuckDB-Erweiterung, die die nahtlose Integration großer Sprachmodelle (LLMs) und von Retrieval-Augmented Generation (RAG) direkt in SQL ermöglicht.

Sie führt MODEL- und PROMPT-Objekte als erstklassige SQL-Entitäten ein, sodass sich LLM-Interaktionen einfach definieren, verwalten und wiederverwenden lassen. Kernfunktionen wie llm_complete, llm_filter und llm_rerank ermöglichen Generierung, semantisches Filtern und Ranking — alles aus SQL.

FlockMTL ist für das schnelle Prototyping LLM-basierter Analytik ausgelegt und mit Batching- und Caching-Funktionen für bessere Leistung optimiert.

📄 Weitere Details und Beispiele finden Sie in der FlockMTL-Dokumentation.

Hinweis: FlockMTL ist Teil der laufenden Forschung des Data & AI Systems (DAIS) Laboratory @ Polytechnique Montréal. Die Erweiterung wird aktiv weiterentwickelt, und einzelne Funktionen können sich ändern. Feedback und Beiträge sind willkommen!

Hinzugefügte Funktionen

function_name	function_type	description	comment	examples
llm_complete	scalar	Generates text completions using a specified language model	Requires a defined prompt and model	[SELECT llm_complete({‘model_name’: ‘default’}, {‘prompt_name’: ‘hello-world’});]
llm_filter	scalar	Filters data based on language model evaluations	returning boolean values	[SELECT * FROM data WHERE llm_filter({‘model_name’: ‘default’}, {‘prompt_name’: ‘is_relevant’, ‘context_columns’: [{‘data’: content}]});]
llm_embedding	scalar	Generates embeddings for input text	Useful for semantic similarity tasks	[SELECT llm_embedding({‘model_name’: ‘default’}, {‘context_columns’: [{‘data’: ‘Sample text’}]});]
llm_reduce	aggregate	Aggregates multiple inputs into a single output using a language model	Summarizes or combines multiple rows	[SELECT llm_reduce({‘model_name’: ‘default’}, {‘prompt_name’: ‘summarize’, ‘context_columns’: [{‘data’: content}]}) FROM documents;]
llm_rerank	aggregate	Reorders query results based on relevance scores from a language model	Enhances result relevance in search applications	[SELECT llm_rerank({‘model_name’: ‘default’}, {‘prompt_name’: ‘rank_relevance’, ‘context_columns’: [{‘data’: content}]}) FROM search_results;]
llm_first	aggregate	Selects the top-ranked result after reranking	Retrieves the most relevant item	[SELECT llm_first({‘model_name’: ‘default’}, {‘prompt_name’: ‘rank_relevance’, ‘context_columns’: [{‘data’: content}]}) FROM search_results;]
llm_last	aggregate	Selects the bottom-ranked result after reranking	Retrieves the least relevant item	[SELECT llm_last({‘model_name’: ‘default’}, {‘prompt_name’: ‘rank_relevance’, ‘context_columns’: [{‘data’: content}]}) FROM search_results;]
fusion_rrf	scalar	Implements Reciprocal Rank Fusion (RRF) to combine rankings	Combines rankings from multiple scoring systems	[SELECT fusion_rrf(score1, score2) FROM combined_scores;]
fusion_combsum	scalar	Sums normalized scores from different scoring systems	Useful for aggregating scores from various models	[SELECT fusion_combsum(score1, score2) FROM combined_scores;]
fusion_combmnz	scalar	Sums normalized scores and multiplies by the hit count	Enhances the impact of frequently occurring items	[SELECT fusion_combmnz(score1, score2) FROM combined_scores;]
fusion_combmed	scalar	Computes the median of normalized scores	Reduces the effect of outliers in combined scores	[SELECT fusion_combmed(score1, score2) FROM combined_scores;]
fusion_combanz	scalar	Calculates the average of normalized scores	Provides a balanced aggregation of scores	[SELECT fusion_combanz(score1, score2) FROM combined_scores;]