v1.6.3 — Better Temporal Search & Retrieval Accuracy
Release Date: 2026-10-07
Changes: v1.6.2 → v1.6.3
Pull Requests: #4116, #5225, #5226, #5249, #5275, #5277, #5309, #5320, #5361, #5428, #5434
Summary
This release sharpens Cognee's search and ingestion reliability: temporal queries are more precise and inclusive, natural-language searches return matching rows and skip irrelevant turn analysis, and ingestion validators prevent malformed data. Several operational fixes improve container builds and external integrations.
Highlights
- Temporal search windows now include their boundary dates and rank date matches more accurately.
- Natural-language searches return matching database rows in the response payload and skip 'turn analysis' to avoid unrelated context.
- Stronger ingestion validation rejects malformed node sets and fixes Notion block rendering (no depth caps).
- Agents restrictions tightened: agents can no longer create nested agents; parent/agent dataset access behavior corrected.
- Operational fixes: Docker images now include git and a DLT dependency is pinned to avoid import failures.
Breaking Changes
- DLT dependency pinned below 1.31: The DLT package used for some ingestion paths is now pinned to versions older than 1.31 to avoid import errors. Action: if you manage your own environment, pin DLT to <1.31 or update your dependency constraints until a newer Cognee release lifts this pin.
- Removed temporal_cognify option: The temporal_cognify configuration option was removed. Action: remove any temporal_cognify references from your configuration or code.
- create_agent refuses nested agents and agents cannot create agents: Nested agents are no longer allowed and once created, agents cannot create other agents. Action: update any automation or scripts that relied on creating nested agents.
- Public sign-up no longer accepts a parent: The signup flow will reject attempts to set a parent account during public sign-ups. Action: remove parent-setting from public signup flows or handle linking after signup.
New Features
- Natural-language search returns row matches in the payload: When you run a natural-language search (a search using plain sentences), Cognee now includes any matching table rows in the response payload. This makes it easier to programmatically access the exact rows that matched the query instead of only high-level text answers.
- Inclusive temporal windows: 'since X' and 'until X' queries now include X in the query window. If you ask for events 'since 2023-01-01' or 'until 2023-01-31', results on those exact boundary dates are now included, giving you predictable behavior for date ranges.
- Notion ingestion removes depth caps and avoids recursive rendering: Notion pages and blocks are now rendered without an artificial depth limit and the Notion renderer avoids recursive expansion. That means deeper or more complex Notion pages ingest more faithfully and without runaway recursion.
- DLT ingestion options grouped as dlt_config and temporal_columns restored: The ingestion configuration for DLT sources (a data-transfer tool used to bring tabular data in) is better organized under a single dlt_config section, and the option to select temporal columns (date/time columns) is restored so you can control which columns feed time-sensitive searches.
Improvements
- Temporal ranking and candidate widening: Date anchors (pieces of content tagged with dates) are ranked by how tightly they match your time window, and Cognee widens the pool of temporal candidates so relevant items near the window edges are more likely to be found. Practically, this makes time-based queries return more relevant items and rank the closest matches higher.
- Natural-language search skips turn analysis: Natural-language searches no longer run 'turn analysis' (analysis that tries to stitch together conversational turns), which reduces noise from unrelated conversational context and leads to cleaner, more relevant results.
- Cleaner logs and privacy: The internal 'extraction reason' (why a piece of content was extracted) is removed from logs to keep logs focused and avoid exposing internal reasoning steps.
- Stronger node_set handling when ingesting rows: The system now refuses malformed node_set values (for example, where a row's node_set is not a JSON list) and rejects bad entries inside node_set lists. This prevents bad ingestion data from corrupting downstream indexing or visualizations.
- Agent and dataset access behavior clarified: Parent accounts correctly reach their agents' datasets, and listing/lookup of agents now correctly limits results to a user's own agents. This makes permissions and dataset visibility more consistent.
Performance
- Search candidate selection widened for temporal queries: Broader candidate selection improves recall for time-sensitive searches, especially when relevant items sit just outside a strict window. This may slightly increase search work for higher recall but yields more useful results.
Security
- Extraction reasons removed from logs to avoid leaking internal extraction rationale.
- Sign-up protection: public sign-up no longer accepts a 'parent' account during creation, preventing accidental or unauthorized linking of accounts at signup.
Bug Fixes
- Return natural-language search rows through payload: Fixed an issue where natural-language queries did not include matching rows in responses.
- Make 'until X' and 'since X' include X: Fixed off-by-one behavior in temporal windows so boundary dates are included.
- Rank dates relative to the time window and log missing windows: Fixed ranking so dates are scored relative to the requested window and added a log when a time window is missing to aid debugging.
- Render Notion blocks without recursive expansion and remove depth caps: Fixed broken or incomplete Notion ingestion where deep nesting could be truncated or cause recursion.
- Refuse malformed node_set and validate entries: Fixed multiple bugs that allowed malformed node_set values or invalid entries — ingestion now fails loudly for malformed input so issues are caught early.
- Stop agents from creating other agents and refuse nested agents at creation: Fixed agent creation so agents cannot create further agents and top-level agent creation enforces non-nested structure.
- Stop warning when a chunk has no summary: Removed noisy warnings when content chunks legitimately lack a summary.
- Make Chinese queries trigger the skill gate: Fixed the recall gate so it fires correctly on Chinese-language queries.
- Install git in Docker images and keep Cypher search raw: Docker images now have git available (fixes build and runtime scripts) and Cypher-based search results retain their raw form where required.
- Pin dlt below 1.31 to keep importable: Avoids a runtime import error by restricting the DLT dependency to a compatible range.
- Various CI and test fixes: Multiple workflow and test updates to keep CI stable (label workflow, PR message requirement, test cleanups).
- Better error handling across pipelines and adapters: Several fixes to avoid swallowed type errors and preserve installer error messages.
Technical Changes
- Large refactors and new adapters: Significant internal refactors and new/updated adapters for Turso (a hosted database), LanceDB and Lancedb-related processes, and vector store components. These are internal plumbing changes to improve scalability and maintainability.
- Pipelines and hooks: After-run hooks in pipeline code now propagate errors instead of silently swallowing them; pipeline task execution code was refactored for clarity and robustness.
- Telemetry and CI adjustments: Updated telemetry scripts and many GitHub workflow files to improve CI reliability and telemetry aggregation.
- Visualizer and graph internals: Node set handling and visualizer stages were refactored to keep node sets out of specific visualizer stages and to simplify marker usage.
- Misc internal: new timestamp parsing utilities, temporal hint helpers, and improved recovery and estimator code for the cognify process.
Dependency Updates
Added:
- dateparser >=1.2,<2
Updated:
- dlt: >=1.9.0,<2 → >=1.9.0,<1.31
- lancedb: >=0.24.3,<1.0.0 → >=0.39.0,<0.40.0
- pylance: >=0.22.0,<=0.36.0 → >=12.0.0,<12.1.0
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 | >=0.39.0,<0.40.0
|
| ladybug | ==0.19.0
|
— The Cognee Team · 2026-10-07