ProductFeaturesDocs
Built InUI ComponentsAdmin ConsoleDatabaseBackendStorageAIRAGAuthSecurity
CompanyStoryAboutPricingBlogRoadmap
Install FLIN
Built into the language

RAG & Semantic Search

No Pinecone. No LangChain. No Chroma. Mark a field as semantic text, save an entity, and FLIN generates embeddings automatically. Search by meaning with one keyword. Build RAG pipelines in a single page.

How Semantic Search Works

One keyword turns any text field into an AI-searchable vector index. No configuration, no external database, no embedding API calls.

1. Mark the Field

Add semantic before any text field. On every save, FLIN generates vector embeddings and indexes them automatically.

2. Search by Meaning

search "query" in Entity finds records by semantic similarity — not keyword matching. "How do plants make food?" finds "Photosynthesis Explained".

3. Build RAG

Search finds relevant documents, injects them as context into a system prompt, and streams an AI answer. The full RAG pipeline in native FLIN code.

From Data to AI Search in 10 Lines

app/knowledge.flin
// 1. Define an entity with semantic text
entity Document {
    title: text @required
    content: semantic text         // ← Embeddings auto-generated on save
    category: text = "general"
}

// 2. Save documents — embeddings generated automatically
doc = Document { title: "Photosynthesis", content: "Plants convert sunlight..." }
save doc

// 3. Search by meaning — not keywords
results = search "how do plants make food from sunlight" in Document

// 4. Or use keyword search for exact matches
exact = keyword_search("photosynthesis", "Document", "content", 10)

// 5. Use results in templates
{for doc in results}
    <div>{doc.title} — {doc.category}</div>
{/for}

Semantic vs. Keyword Search

Keyword search finds exact words. Semantic search finds meaning. FLIN gives you both — side by side, in the same query language.

Semantic Search
// Finds by MEANING
search "baking desserts at home"
    in Document

// Matches: "Chocolate Chip Cookie
// Recipe" even though "baking" and
// "desserts" don't appear in content

// Uses vector embeddings
// Cosine similarity ranking
// Works across languages
Keyword Search
// Finds by EXACT WORDS
keyword_search(
    "error E-4021",
    "Document", "content", 10
)

// Matches: documents containing
// the exact string "E-4021"

// Uses BM25 ranking
// Multilingual stop words
// Best for codes & identifiers

Semantic Wins

"Tips for better rest" finds "Healthy Sleep Habits". "Ancient kingdoms in Africa" finds "History of West Africa". Meaning over words.

Keyword Wins

"Error E-4021" finds the exact error code. "PostgreSQL max_connections" finds the config parameter. Precision over meaning.

Use Both

FLIN automatically falls back to keyword search when the vector store is unavailable. Best of both worlds with zero configuration.

5 Embedding Providers. One Config Line.

Choose your embedding provider in flin.config. Switch between local and cloud without changing any application code.

Local (FastEmbed)

Free, offline, 384-dim. Model downloads automatically (~100 MB).

embeddings: "local"

OpenAI

text-embedding-3-small, 1536-dim. Best quality for English.

embeddings: "openai"

OpenAI Vector Store

Hosted vector DB. Auto-chunking, persistent, scalable to millions.

embeddings: "openai-vectorstore"

Voyage AI

voyage-3, 1024-dim. Optimized for retrieval quality.

embeddings: "voyage"

Cohere

embed-multilingual-v3.0, 1024-dim. Best for multilingual.

embeddings: "cohere"
flin.config
# Local embeddings — free, offline, zero API keys
ai {
    embeddings: "local"
}

# OpenAI embeddings — best quality
ai {
    embeddings: "openai"
    embedding_model: "text-embedding-3-small"
}

# OpenAI Vector Store — hosted, persistent, scalable
ai {
    embeddings: "openai-vectorstore"
}

# Cohere — best for multilingual content
ai {
    embeddings: "cohere"
    embedding_model: "embed-multilingual-v3.0"
}

RAG Without the Plumbing

In other frameworks, RAG means stitching together a vector database, an embedding API, a chunking strategy, and an LLM client. In FLIN, it is a page.

RAG Pipeline in FLIN
// 1. User asks a question
question = "What is the vacation policy?"

// 2. Find the 5 most relevant documents
docs = search question in Document by content limit 5

// 3. Build context from search results
fn buildContext() {
    context = ""
    for doc in docs {
        context = context + "--- " + doc.title + " ---\n" + doc.content + "\n\n"
    }
    return context
}

// 4. Ask the AI with document context
answer = ask_ai(question, [
    "provider": "claude",
    "system_prompt": "Answer based on these documents:\n" + buildContext()
])

Search

Semantic search finds the 5 most relevant documents from your knowledge base. Meaning-based, not keyword-based.

Context

Found documents are injected into the system prompt as context. The LLM answers based on your data, not hallucinations.

Answer

ask_ai() sends the question with context to any of 8 LLM providers. Stream the response or get it in full.

The Full Document Pipeline

Upload a PDF. FLIN extracts the text, generates embeddings, indexes them, and makes the content searchable — all on save.

Upload → Parse → Index → Search
// Upload and parse a document
route POST /upload {
    validate {
        document: file @required @document
        title: text @required
    }

    // Extract text from PDF, DOCX, CSV, etc.
    content = document_extract(body.document)

    // Save — embeddings generated automatically
    doc = Document { title: body.title, content: content }
    save doc

    response.created(doc)
}

// Search across all uploaded documents
results = search "quarterly revenue projections" in Document

Upload

Multipart file upload with validators: @max_size("50MB"), @extension(".pdf", ".docx"), @document

Parse

document_extract() handles 9 formats: PDF, DOCX, HTML, CSV, XLSX, JSON, YAML, RTF, XML

Embed

Vector embeddings generated on save. 5 providers: local (free), OpenAI, Voyage, Cohere, OpenAI Vector Store

Search

Cosine similarity ranking with automatic BM25 keyword fallback. Results sorted by relevance score

Multilingual by Default

Keyword search uses 220+ stop words across 5 languages. Semantic search with Cohere or FastEmbed works across languages out of the box.

5 Languages

English, French, Spanish, Portuguese, and German stop words built into the keyword search engine. No configuration needed.

Cross-Lingual Search

With multilingual embedding models (FastEmbed, Cohere), search in English across French documents — or any combination.

Smart Filtering

Stop words are filtered before embedding and keyword indexing. "Le", "les", "des", "un" in French; "der", "die", "das" in German — all handled.

Build AI-Powered Search Today

Add semantic text to any field. Upload documents. Search by meaning. Build RAG pipelines. Ship AI features in minutes, not weeks.

curl -fsSL https://flin.sh | bash
Install FLIN All AI Features