NetSendo v2.7.0: Semantic AI Brain & Rich-Text Editor

Your AI assistant is only as good as the context it receives. If you have spent months filling NetSendo Brain with product details, case studies, objection handling and campaign notes, you may have found a frustrating limitation. Until now, Brain added the five best-rated entries from a few fixed categories to every prompt, whatever the task was about. In a large knowledge base, most of what Brain knew never reached it.

NetSendo v2.7.0 fixes this with semantic knowledge search on the GOLD plan. Brain now looks for the knowledge that matches each task by meaning. The release also adds a real rich-text editor to the email template builder and makes Docker installs and updates safer.

🚀 What's New in v2.7.0

Semantic Brain Knowledge Base (GOLD)

Category-based retrieval works like a librarian who hands you the five most popular books from the same shelf every time, whatever you asked about. Semantic search works like a librarian who reads your question first and then picks the right pages from the whole library.

Here is how it works:

• Chunking: every entry is split into chunks of about 1,200 characters.
• Embeddings: each chunk becomes a vector, a list of numbers that represents its meaning.
• Matching: each task is compared with your chunks by meaning. A task can be a chat message, an email or campaign topic, a goal, a playbook brief or a cron task.
• Bonuses: a chunk gets a small boost if it contains words found by the full-text index, belongs to the task's categories or comes from a verified entry. Matches that are much weaker than the best one are left out.

Brand voice and company entries are still included in writing tasks. Entries can now hold up to 100,000 characters on GOLD, and 10,000 on SILVER.

ℹ️ Why no vector database? The VECTOR type in MySQL 9 has no distance function outside HeatWave, and MySQL Community cannot search vectors. NetSendo therefore stores vectors (768 dimensions where the model can shorten them) in a new knowledge_chunks table and compares them in PHP. You don't need to add Qdrant, pgvector or any other container.

The index updates itself. When you add or change an entry, a queued IndexKnowledgeEntry job embeds it, and changed text removes its old vectors. An hourly brain:index-knowledge command embeds anything still missing. Saving an entry never fails because the provider is down. If you switch to a different model, the whole knowledge base is embedded again in the background, and vectors from different models are never mixed. The API endpoint GET /api/v1/brain/knowledge?search= also searches by meaning on GOLD.

Rich-Text Editor for Template Text Blocks

Before this release, the text block field was a plain text box. If you changed one word of AI-generated content, the block lost its links, bold text, headings and colours. It is now a TipTap rich-text editor with bold, italic, underline, links, lists, line breaks (Shift+Enter), and undo and redo.

• Byte-for-byte safe: nothing is saved until you change the text. Inline styles such as #64748b and mso-* properties are kept exactly as written.
• Insert variable at the cursor: variables, snippets and signatures are inserted where your cursor is, not at the end of the text.
• Email-safe HTML: scripts, iframes, forms, <font> tags and CSS classes are removed, and javascript: links are refused. Placeholders such as [[unsubscribe]] work as link addresses. List items are stored without inner <p> tags, so email clients don't add extra margins.

Only One Container Migrates the Database

The app, scheduler, queue and reverb containers run the same entrypoint. Before this release, all four could run migrations at the same time, which caused errors such as Table 'users' already exists and could leave a migration half-applied. Migrations and seeding now run under a MySQL named lock (GET_LOCK). One container prepares the database while the others wait. If that container dies, MySQL releases the lock automatically.

💼 Benefits & Use Cases

🛠️ Step-by-Step Guide

1. Back up your database

   This release includes three migrations: the knowledge_chunks table, two new columns in ai_brain_settings, and knowledge_entries.content changed to MEDIUMTEXT.

2. Update the image

   docker compose pull && docker compose up -d

   The new image already protects migrate and db:seed with a database lock.

3. Update the compose file and entrypoint (recommended)

   Your installation mounts its own copy of the entrypoint, so pulling the image does not replace it. Run git pull, or download docker-compose.yml and docker/php/docker-entrypoint.sh from the v2.7.0 release. With both files, only app migrates the database, and the other containers start once it is done.

4. Configure semantic search

   Go to Brain → Settings → Knowledge base → Semantic search. Choose an integration and a model, or leave it on automatic. NetSendo tests the model before saving it and shows how many entries are indexed.

5. Index existing entries

   The hourly job will index your existing entries, but you can start indexing right away:

   docker compose exec app php artisan brain:index-knowledge --sync

⚠️ Common pitfall: Anthropic and Grok have no embedding endpoint. If those are your only AI integrations, add an OpenAI, Gemini, OpenRouter or Ollama integration in Settings → AI Integrations first. See Integrations.

🔧 Other Fixes and Improvements

• The install.sh script now produces a working installation. It downloads every file the compose file mounts, generates APP_KEY and a random database password, refuses to overwrite an existing .env file, and waits until the app answers.
• The Delete key no longer removes the whole selected block while you type in a text field in the template builder.
• Clicking into a text block no longer strips its formatting.
• Web research entries reach Brain again. Entries with categories the AI made up now move to Insights, and the old category is kept as a tag.
• New knowledge base entries now start with the Company category, so saving without choosing a category no longer fails silently.
• An Ollama integration behind a proxy now sends its optional API key as a bearer token.
• Release titles with build metadata now display correctly, and HTML is converted to text with the browser's parser instead of a regular expression.
• The README, DOCKER_INSTALL, CONTRIBUTING and SECURITY documents have been updated, and broken links to forum.netsendo.com have been removed.

🎉 Conclusion

In v2.7.0, NetSendo Brain retrieves knowledge by meaning, and everything runs on your own MySQL with the AI provider you choose. Template editing keeps your formatting, and Docker updates are more reliable. You keep ownership of your data and don't depend on a single vendor.