Opensolr Changelog

The LangChain, LlamaIndex, Haystack, MCP and Laravel Scout packages are out at 0.3.0, and every one of them is now covered end to end by a live test suite that runs against the real API. Three things that suite found are fixed in this release: a mistyped hybrid search mode returned a silently different result set instead of an error, argument checks ran after the embedding call rather than before it, and a dropped connection surfaced as a raw network error instead of being retried.

api.opensolr.com now serves only the AI and Data Ingestion endpoints. Embeddings, hybrid search, ai_summary and the ingestion queue answer there; every management endpoint (indexes, configuration, crawler, backups, security) is served exclusively by opensolr.com and answers 403 on the AI host. No client change is needed — all official integrations already use the right host for each call.

The API Reference page now tells the truth about errors and hosts. It explains which endpoints live on opensolr.com (management) and which on api.opensolr.com (AI & Data Ingestion), lists the four HMAC-signed endpoints, and documents the real error responses (HTTP 200 with "status": false for application errors, 403 for authentication, 429 with Retry-After). Two commands that referenced non-existent endpoints were removed from the Solr Migration guide.

All AI integrations updated. LangChain, LlamaIndex, Haystack, the MCP server and Laravel Scout all ship the new answer behaviour. Update with pip install -U or composer update.

The ai_summary API is now fully deterministic: your instruction is the prompt. Whatever you send in instruction is exactly what the model reads, byte for byte. Nothing reorders it, relabels it, appends to it or shortens it, so a prompt you tuned behaves the same on every call. Retrieval stays yours: run embed_and_search, take your top results and place them in the prompt where you want them. The query and context parameters still work for existing integrations. Prompts up to 100,000 characters are accepted.

Every AI integration ships the improved answer behaviour. LangChain, LlamaIndex, Haystack, the MCP server and Laravel Scout all now send a grounded answer instruction by default and drop weakly matching documents before generating, so RAG answers are direct and stay on topic. Update with pip install -U or composer update.

embed_and_search now matches your hosted search page exactly. The one-shot hybrid search endpoint runs the platform’s full tuned pipeline: your index’s saved Search Tuning applies automatically, natural-language questions are processed the same way as on your search page (stopwords cleaned for keyword matching, while the full phrase goes to the semantic embedder), and every tuning knob (fw_*, mm, search_mode, vector_topk, quality_boost, and more) accepts per-request overrides. Results returned via the API are now identical to your search page. Full parameter reference in the AI-Hints API guide.

Your Search Tuning now applies everywhere — including all AI integrations. Everything you configure in Index Settings → Search Tuning (semantic↔lexical balance, field weights, minimum match, search mode, vector candidate pool, content quality boost) now applies automatically to every search and every RAG answer made through the LangChain, LlamaIndex, Haystack, Laravel Scout, and MCP integrations — the same tuned pipeline as your hosted search page. Each knob can also be overridden per call from code via the new tuning parameter. Packages updated: LangChain, LlamaIndex, Haystack (v0.2.5), Laravel Scout (v0.2.5), MCP (v0.2.6).

All AI integrations upgraded to v0.2.0. Writes in every integration — LangChain, MCP, LlamaIndex, Haystack, Laravel Scout — now go through the Data Ingestion API: documents are queued and enriched server-side (embeddings, sentiment, language, and all derived fields — identical to the Web Crawler), with progress visible in your Control Panel. Every integration also gains a lexical-only mode: pure keyword search with zero AI quota that works on any Opensolr index, including non-vector ones. Plus delete-by-query support and server-side PDF/DOCX text extraction via rtf:true.

New is_query parameter on the embed and batch embed API endpoints. Set is_query=1 when embedding search queries to get optimized retrieval vectors.

The Data Ingestion API now returns a doc_ids array in every successful response, showing the auto-generated document ID (md5 of uri) for each document in your batch. Use these IDs to track, query, or update specific documents in your index.

URI is now mandatory for every document in the Data Ingestion API. The document ID is always generated as md5(uri), making the URI the single source of truth for document identity. Same URI = same document. Resubmitting a URI updates the existing document. Duplicate URIs in pending jobs are automatically rejected to prevent accidental double-indexing.