@memberjunction/content-autotagging

AI-powered content ingestion, autotagging, and vectorization engine for MemberJunction. Scans content from multiple sources (local files, websites, RSS feeds, cloud storage), extracts text from documents, uses LLMs to generate weighted tags and metadata attributes, and vectorizes content for semantic search.

Read these guides first if you're working on tag classification, taxonomy growth, or governance:

• Content Autotagging Guide — pipeline architecture, prompt structure, source-type providers
• Taxonomy & Tagging Guide — the tag taxonomy itself: 4+1-tier resolver, per-tag governance, scoping, the suggestion queue, Tag Health, and per-source configuration knobs

Overview

The @memberjunction/content-autotagging package provides an extensible framework for ingesting content from diverse sources and leveraging AI models to extract meaningful tags, summaries, and metadata. Built on the MemberJunction platform, it helps organizations automatically organize and categorize their content. The engine uses the managed "Content Autotagging" AI prompt via AIPromptRunner (rather than direct BaseLLM calls), enabling prompt versioning, model routing, and centralized prompt management.

graph TD
    A["AutotagBaseEngine
(Orchestrator)"] --> B["Content Sources"]
    B --> C["Local File System"]
    B --> D["Websites"]
    B --> E["RSS Feeds"]
    B --> F["Cloud Storage
(Azure Blob)"]

    A --> G["Text Extraction"]
    G --> H["PDF Parser"]
    G --> I["Office Parser"]
    G --> J["HTML Parser
(Cheerio)"]

    A --> K["AIPromptRunner
(Content Autotagging prompt)"]
    K --> L["Tag Generation
(with weights)"]
    K --> M["Attribute Extraction"]

    A --> V["Vectorization"]
    V --> W["Embedding Model"]
    V --> X["Vector DB Upsert"]

    A --> N["Content Items
(Database)"]
    A --> O["Content Item Attributes
(Database)"]

    style A fill:#2d6a9f,stroke:#1a4971,color:#fff
    style B fill:#7c5295,stroke:#563a6b,color:#fff
    style C fill:#2d8659,stroke:#1a5c3a,color:#fff
    style D fill:#2d8659,stroke:#1a5c3a,color:#fff
    style E fill:#2d8659,stroke:#1a5c3a,color:#fff
    style F fill:#2d8659,stroke:#1a5c3a,color:#fff
    style G fill:#b8762f,stroke:#8a5722,color:#fff
    style K fill:#7c5295,stroke:#563a6b,color:#fff
    style V fill:#2d6a9f,stroke:#1a4971,color:#fff
    style N fill:#2d6a9f,stroke:#1a4971,color:#fff
    style O fill:#2d6a9f,stroke:#1a4971,color:#fff
Copy

Key Features

• AIPromptRunner integration: Uses the managed "Content Autotagging" prompt, enabling prompt versioning and model routing through MJ's prompt management system (no direct BaseLLM calls)
• Tag weights: Each generated tag includes a relevance weight (0.0--1.0) indicating how strongly the tag relates to the content
• Streaming crawl → classify pipeline: Content items flow into the LLM batcher as the crawler discovers them — total wall-clock is roughly max(crawl, classify) instead of crawl + classify. Backwards-compatible with array callers
• Batch processing: Configurable batch size (default: 20) with concurrent processing within each batch
• Parallel tagging + vectorization: Tagging and vectorization run in parallel for maximum throughput
• Per-source/type embedding model selection: Cascade resolution for embedding model and vector index -- source override, then content type default, then global fallback (first active vector index)
• Real-time progress reporting: AutotagProgressCallback provides per-item progress updates during processing
• Graceful provider skip: Providers skip gracefully when no content sources are configured for their type
• Per-source run budgets: MaxItemsPerRun, MaxNewTagsPerRun, MaxTokensPerRun, MaxCostPerRun, MaxNewTagsPerItem cap each run's resource use. When a cap trips, the run pauses gracefully — the next invocation picks up where it left off (change-detection skips already-processed items)
• Content-stable change detection: Hash is computed over extracted body text, not raw HTML, so incidental markup changes (timestamps, CSRF tokens, build hashes) don't trigger spurious re-classification. Single HTTP fetch derives both the hash and the page text
• Website crawler defaults that work out of the box: MaxDepth=2, recursive crawl enabled, conservative URL normalization (fragment-strip, trailing-slash collapse, query-param sort) so common URL variants dedupe correctly

Installation

npm install @memberjunction/content-autotagging
Copy

Content Processing Pipeline

sequenceDiagram
    participant Source as Content Source
(crawler/streamer)
    participant Engine as AutotagBaseEngine
    participant Extract as Text Extractor
    participant Prompt as AIPromptRunner
    participant Vec as Embedding + VectorDB
    participant DB as Database

    loop For each discovered item (streaming)
        Source->>Source: Fetch URL/file (1 HTTP request)
        Source->>Source: Hash extracted body text
        Source->>DB: Lookup existing checksum (scoped to source)
        alt Unchanged
            Source-->>Source: Skip — no LLM cost
        else New / changed
            Source->>Engine: Yield ContentItem
        end
    end
    Engine->>Engine: Accumulate items into BatchSize chunks
    Engine->>Engine: Apply rate limit + budget gate
    Engine->>Extract: Extract text (PDF/Office/HTML)
    par Tagging
        Engine->>Prompt: Run "Content Autotagging" prompt
        Prompt-->>Engine: Tags (with weights) + Attributes
    and Vectorization
        Engine->>Vec: Embed text + upsert to vector DB
        Vec-->>Engine: Vectorization result
    end
    Engine->>DB: Save ContentItem + Tags + Attributes
    Engine->>Engine: Check RunBudget; pause if exhausted
    Engine->>DB: Create / update ProcessRun record (checkpoint)
Copy

The crawl and LLM stages run concurrently against a bounded buffer. The LLM consumer starts processing as soon as the first batch fills — it does not wait for every content source to finish crawling.

Content Sources

All sources extend AutotagBase, which provides the common interface for content discovery and ingestion. Each source's Autotag() method accepts an optional AutotagProgressCallback for real-time progress reporting. Sources skip gracefully when no content sources of their type are configured in the database.

Supported File Formats

Tag Weights

The LLM prompt returns tags with relevance weights between 0.0 and 1.0 indicating how strongly each tag relates to the content. Both old-style (plain string array) and new-style (object with tag + weight) responses are supported:

// New format (preferred) — returned by the "Content Autotagging" prompt
[
  { "tag": "machine learning", "weight": 0.95 },
  { "tag": "neural networks", "weight": 0.82 },
  { "tag": "data science", "weight": 0.70 }
]

// Legacy format — auto-normalized with weight 1.0
["machine learning", "neural networks", "data science"]
Copy

Embedding Model and Vector Index Resolution

The engine resolves the embedding model and vector index for each content item using a three-level cascade:

1. Content Source override: If the source has EmbeddingModelID and VectorIndexID set, those are used
2. Content Type default: If the source has no override, the content type's defaults are used
3. Global fallback: If neither source nor type specifies, the first active vector index in the system is used

Items sharing the same (embeddingModel, vectorIndex) pair are grouped and processed together for efficient batching.

Usage

RSS Feed Processing

import { AutotagRSSFeed } from '@memberjunction/content-autotagging';

const rssTagger = new AutotagRSSFeed();
await rssTagger.Autotag(contextUser, (processed, total, currentItem) => {
    console.log(`[${processed}/${total}] Processing: ${currentItem}`);
});
Copy

Website Content Processing

import { AutotagWebsite } from '@memberjunction/content-autotagging';

const websiteTagger = new AutotagWebsite();
await websiteTagger.Autotag(contextUser);
Copy

Per-source crawl knobs (see Website Crawl Settings below) are read from ContentSourceParam rows; per-source run-budget knobs (see Run Budgets below) are read from the source's Configuration JSON with a ContentSourceParam override.

Local File System Processing

import { AutotagLocalFileSystem } from '@memberjunction/content-autotagging';

const fileTagger = new AutotagLocalFileSystem();
await fileTagger.Autotag(contextUser);
Copy

Azure Blob Storage Processing

import { AutotagAzureBlob } from '@memberjunction/content-autotagging';

const blobTagger = new AutotagAzureBlob(
  process.env.AZURE_STORAGE_CONNECTION_STRING,
  'your-container-name'
);
await blobTagger.Authenticate();
await blobTagger.Autotag(contextUser);
Copy

Direct Engine Usage

import { AutotagBaseEngine } from '@memberjunction/content-autotagging';

const engine = AutotagBaseEngine.Instance;

// Process content items with custom batch size
await engine.ExtractTextAndProcessWithLLM(contentItems, contextUser, batchSize);

// Vectorize content items (runs in parallel with tagging)
const result = await engine.VectorizeContentItems(contentItems, tagMap, contextUser, batchSize);
console.log(`Vectorized: ${result.vectorized}, Skipped: ${result.skipped}`);
Copy

Creating a Custom Content Source

import { AutotagBase, AutotagProgressCallback } from '@memberjunction/content-autotagging';
import { RegisterClass } from '@memberjunction/global';

@RegisterClass(AutotagBase, 'AutotagCustomSource')
export class AutotagCustomSource extends AutotagBase {
  public async SetContentItemsToProcess(contentSources) {
    // Fetch and create content items from your custom source
    return contentItems;
  }

  public async Autotag(contextUser, onProgress?: AutotagProgressCallback) {
    const contentSourceTypeID = await this.engine.setSubclassContentSourceType(
      'Custom Source', contextUser
    );
    const contentSources = await this.engine.getAllContentSources(
      contextUser, contentSourceTypeID
    );
    if (contentSources.length === 0) return; // Skip gracefully
    const contentItems = await this.SetContentItemsToProcess(contentSources);
    await this.engine.ExtractTextAndProcessWithLLM(contentItems, contextUser);
  }
}
Copy

Website Crawl Settings

The AutotagWebsite provider reads its crawl behavior from two layered sources:

1. Typed Configuration.Website sub-object on the content source (canonical for new sources, set via the Content Source form's "Website Crawler Settings" section).
2. ContentSourceParam rows with matching keys (legacy storage / per-instance sharper override — wins over the typed sub-object when both are present).

Each knob has a sensible default so newly-created sources work without configuration; override only the ones you need.

URL normalization (always on). Before any visited-set check, URLs are normalized: fragments (#section) stripped, single trailing slash collapsed (except root /), query parameters sorted. Path case is preserved (RFC 3986). This catches the common variants (/page vs /page/, ?a=1&b=2 vs ?b=2&a=1) without merging genuinely distinct case-sensitive paths.

Change detection. Each discovered URL is fetched once. The SHA-256 hash is computed over the extracted body text (not raw HTML), so incidental markup churn — server-rendered timestamps, CSRF tokens, build hashes, ad rotators — doesn't trigger spurious re-classification. The hash query is scoped to the current ContentSourceID so 404 boilerplate on one site can't silently skip a real page on another.

Run Budgets

Every Autotag run can be capped along several axes. Caps are checked after each batch; when any cap exhausts, the run pauses gracefully via the existing CancellationRequested machinery. Next invocation picks up where it left off — the change-detection short-circuit (above) skips already-processed items, so re-crawl wastes only HTTP, not LLM cycles.

Priority order when several caps trip in the same batch: items → tags → tokens → cost. The items knob ranks first because it's the most user-intuitive ("do at most 100 today, do the rest tomorrow") and not tied to a specific model's pricing or tokenization.

MaxItemsPerRun / MaxTokensPerRun / MaxCostPerRun may also be overridden per-source-instance via a matching ContentSourceParam row; that override beats the ConfigurationObject default. Tag-related knobs come from ConfigurationObject only.

Note on resume granularity: today the LLM phase's LastProcessedOffset cursor is array-indexed and works fully for callers that materialize an array (e.g. AutotagEntity). The streaming path relies on change-detection dedup for "functional resume" — pages already processed are skipped on re-crawl, so the work picks up correctly even though the HTTP discovery happens again. True crawl-side resume (persisted discovered-URL list) is a planned follow-up.

Database Entities

Dependencies

License

ISC