Graphlit

Graphlit is a managed context layer for ingesting, enriching, searching, and retrieving unstructured content. Its pricing page lists PDFs, audio, video, web pages, feeds, Notion, blob storage, text and vector search, RAG conversations, workflows, and MCP access. However, the same official page now says Graphlit is winding down, new signups are closed, and only existing accounts remain available during the transition. That status dominates the review: new buyers should not begin an implementation. Existing customers should inventory feeds, items, embeddings, workflows, API dependencies, MCP clients, and chatbots. Export source content and derived metadata, document schemas and identifiers, choose a replacement, and run it in parallel against a fixed retrieval test set.

Official pricing checked September 18, 2026 lists Free at $0; Hobby at $49 per month plus $0.10 per credit, up to 10 GB and 10,000 items; Starter at $199 plus $0.09 per credit, up to 100 GB and 100,000 items; and Growth at $999 plus $0.08 per credit with unlimited storage and items listed. These figures may inform transition billing for existing customers, but signup is closed. The honest use case is migration and continuity planning, not new adoption.

Relevant comparisons include LlamaIndex, Unstructured, Pinecone, Qdrant, RAGFlow. These products cover different layers, so compare the same representative task and include setup and operating labor rather than ranking feature counts.

Run a bounded pilot before production. Choose a representative workflow, use non-sensitive data, and define success in terms of completion rate, accuracy, human correction time, latency, retries, and total cost. Test authentication expiry, rate limits, malformed input, dependency outages, cancellation, and recovery. Start with read-only permissions and least-privilege credentials. Require human approval before messages, purchases, deletions, deployments, remediation, or other consequential writes. Review retention, deletion, audit logs, subprocessors, regional hosting, support, and export paths. If the system uses a browser or third-party service, confirm that the workflow complies with account permissions, privacy duties, and site terms. A polished demonstration is not evidence of dependable production performance. Keep the current process available during the trial, inspect failures instead of averaging them away, and document an exit plan before the tool becomes a dependency. Assign one owner to the pilot and record configuration, permissions, test cases, observed failures, and unresolved assumptions. Calculate cost for a quiet month, an expected month, and a high-volume month, including seats, usage, model fees, storage, network traffic, engineering time, and support. Verify that data and configuration can be exported in a usable format. Recheck current documentation and pricing immediately before publication or procurement because SaaS capabilities and commercial terms change. The strongest buying decision links one recurring problem to a measurable improvement and identifies who will monitor the system after launch; it does not rely on a broad claim that AI will automate the workflow.