v1.6.0 — Keyless workflows & pipeline reliability

Release Date: 2026-09-18 Changes: v1.5.4rc1 → v1.6.0

Pull Requests: [#4920], [#4994], [#5106]

---

Summary

This release makes running Cognee without a cloud LLM key more usable and predictable, strengthens pipeline startup and recovery behaviors, and finishes wiring image/audio/embedding models end-to-end. It also introduces developer-facing tooling for integrations, tightens telemetry privacy, and fixes numerous bugs across ingestion, recall, and the CLI.

Highlights

• Keyless-first experience: local model downloads are announced and many flows now run cleanly with no LLM key.
• Pipeline robustness: pipeline runs are stamped at start and recover reliably after crashes, preserving completed documents.
• Model wiring: image and audio transcription models are passed through end-to-end and the embedding model is recorded per dataset to ensure consistent embeddings.
• Telemetry & privacy: dataset names are fingerprinted before leaving the host and error telemetry now reports full routes for better diagnostics.
• Developer tooling: a new cognee-mcp client/server and CLI inclusions for Enola improve integration and testing workflows.

Breaking Changes

• GLiNER removed from default Docker image — The default Docker image no longer bundles GLiNER tooling. If you relied on GLiNER being present in the standard image, you must add it to your image or install it separately.
• Database & adapter refactors may require review — Internal database and vector-store adapters were refactored (including changes to Postgres/hybrid adapters). Self-hosted deployments or custom adapter integrations should review their configuration and custom adapter code for compatibility.

New Features

• Keyless first-use downloads and notices — When you run Cognee without a cloud LLM key, the system will now announce when it needs to download local model files and explain what is happening. This helps first-time users understand local model activity and avoids silent/blocking downloads.
• IMAGE_TRANSCRIBE_MODEL wiring — Image transcription (turning images into text) is now passed through to the LLM adapters responsible for image processing. This makes image-based uploads and transcriptions behave consistently end-to-end.
• Audio and image LLM support through legacy adapters — Audio and image model selection are now wired through legacy adapter layers, so audio transcription and image tasks will use the configured models consistently across adapters.
• MCP client/server (cognee-mcp) added — A new Python client/server package has been added to support improved external integrations and end-to-end tests for recall/summary workflows. This toolset helps integrators interact programmatically with Cognee's recall/collation features.
• Record embedding model per dataset — The embedding model used to create vectors is now recorded on each dataset (a dataset is a collection of documents you add). Cognee checks that the same model is used when reusing or updating embeddings, preventing silent mismatches that could degrade search quality.

Improvements

• Better keyless stage handling — Pipeline stages that require an LLM will now skip cleanly when no LLM key is present, preventing partial failures and confusing errors during keyless runs.
• Skip keyless trace context extraction when keyless — The session context extraction no longer tries to extract LLM-related trace data when running without a key, reducing noise and errors in logs.
• Stop probing the LLM during add() — Adding documents no longer probes the cloud LLM to detect availability, improving responsiveness and preventing unnecessary LLM requests during ingestion.
• Pipeline origin stamping at start — Pipeline runs are stamped with their origin at start time so subsequent diagnostics and recovery steps can accurately attribute run activity.
• Fail search/recall on unresolvable dataset names — If a dataset name cannot be resolved, search and recall now fail clearly instead of proceeding silently with incorrect assumptions.
• Scope agent prompt recall bodies to a dataset — Agent prompts that ask the system to recall content are now scoped explicitly to the requested dataset, reducing cross-dataset leakage in recall results.
• Telemetry: pass dataset ids & mark fingerprints — Telemetry events now include dataset IDs (internal identifiers) and mark names that were fingerprinted, improving observability while reducing raw name leakage.

Performance

• More reliable pipeline recovery — Startup recovery now preserves documents completed by runs that were abandoned, reducing duplicated work and improving recovery speed after crashes.
• Reduced unnecessary LLM probing — By avoiding LLM probes during add() and keyless runs, ingestion and local flows are faster and less dependent on external LLM latency.

Security

• Fingerprint dataset names for telemetry — Human-readable dataset names are hashed (fingerprinted) before telemetry leaves the host. This reduces privacy risk when telemetry is collected while preserving the ability to correlate events.
• Error telemetry includes full route context — When errors are reported, the full API route is now included in telemetry to make debugging faster while keeping sensitive names protected by fingerprinting.

Bug Fixes

• Align dataset permission denial contract — API permission denial behavior was corrected so clients get consistent and explicit permission errors.
• Fix session bridge and session-layer refactors so sessions run without an LLM key — Multiple fixes ensure session-related features no longer require a cloud LLM key to operate.
• Record and check embedding model once per dataset — Avoids repeated and inconsistent embedding-model checks and prevents embedding/recall mismatches.
• Don't show LLM key error when GLiNER is used for file uploads — Upload flows that use alternate upload tooling no longer spuriously report missing LLM key errors.
• Let route stamp follow content on refresh — When content is refreshed, the route-origin stamp follows the updated content so telemetry and audit trails remain accurate.
• Keep graph results when a session entry leads in the CLI — CLI recall now preserves graph results if a session entry is the leading item.
• Wire image and audio models through adapters — Fixed previous gaps so image and audio tasks use configured models consistently.
• Fix various test flakiness and CI issues — Many tests and CI workflows were pinned or fixed (Enola CLI, nested release tests, platform lookups) to stabilize releases.
• Fix recall routing and search-type span calls — Retrieval/routing logic and span search-type calls were normalized and several edge-case bugs were resolved.
• Fail search/recall on unresolvable dataset names — Ensures clearer errors and avoids unexpected results when dataset names are invalid.

Technical Changes

• Large refactor of session and prompt handling — The session layer has been consolidated and moved into the user prompt representation; user/system prompts were split and several session persistence modules were reorganized.
• Major 'improve' subsystem additions — New modules added for staged improvement, result collapsing, gating, and graph-change recording to support richer automated improve workflows.
• Retrieval and search rework — Retrieval code paths were reorganized (completion retriever, graph-completion retrievers, hybrid retriever), and a clearer RecallScope was added for session-aware recall.
• New background tasks and pipeline operations — Pipeline run lifecycle (start, logging, error handling, stamping) and background tasks were refactored to make run recovery and diagnostics more robust.
• New testing and examples for MCP and recall summaries — Additional tests and example fixtures were added to cover the new mcp tooling and recall/summary behaviors.

Dependency Updates

Added:

• enola-cli ==0.4.21
• fastembed <=0.8.0
• onnxruntime >=1.24.1

Updated:

• pydantic-settings: >=2.2.1,!=2.12.,!=2.13.,!=2.14.0,!=2.14.1,<3 → >=2.7.0,!=2.12.,!=2.13.,!=2.14.0,!=2.14.1,<3

Compatibility

Component	Supported / Required
Python	>=3.10,<3.15
pydantic	>=2.10.5
litellm	>=1.83.7,<1.97.0
fastapi	>=0.116.2,<1.0.0
sqlalchemy	>=2.0.39,<3.0.0
lancedb	>=0.24.3,<1.0.0
ladybug	==0.19.0

---

— The Cognee Team · 2026-09-18