[PR #403] [MERGED] Fix txtai IndexIDMap 'nprobe' fallback loop by skipping incompatible FAISS index load #707

New issue

Closed

opened 2026-03-13 21:05:04 +03:00 by kerem · 0 comments

kerem commented

2026-03-13 21:05:04 +03:00

Owner

📋 Pull Request Information

Original PR: https://github.com/AJaySi/ALwrity/pull/403
Author: @AJaySi
Created: 3/9/2026
Status: ✅ Merged
Merged: 3/9/2026
Merged by: @AJaySi

Base: main ← Head: codex/fix-nprobe-attribute-error-in-search

📝 Commits (2)

cd8582e Fix txtai nprobe fallback to avoid reloading incompatible faiss index
952824a Stabilize txtai nprobe handling without dropping loaded index state

📊 Changes

1 file changed (+70 additions, -47 deletions)

View changed files

📝 backend/services/intelligence/txtai_service.py (+70 -47)

📄 Description

Motivation

Backend logs showed repeated 'IndexIDMap' object has no attribute 'nprobe' errors when the service attempted to fallback from FAISS to numpy but then reloaded a persisted FAISS ANN index, causing the same failure to recur.
The goal is to make TxtaiIntelligenceService more resilient by ensuring backend switches do not reintroduce incompatible persisted ANN state and by providing safe retry/fallback code paths for search, similarity, and clustering.

Description

Changed _initialize_embeddings to accept load_existing_index: bool and added logic to skip loading a persisted index when instructed (prevents reloading incompatible FAISS state).
Added _is_nprobe_incompatibility to centralize detection of the FAISS IndexIDMap/nprobe error pattern and _switch_to_numpy_backend to encapsulate backend switching and reinitialization behavior.
Hardened search to retry against a numpy backend with load_existing_index=False and, if a retry still raises the same error, force a non-ANN retrieval path by calling search(..., index=False).
Hardened get_similarity to switch to numpy on the incompatibility and, if similarity() keeps failing, fall back to a cosine similarity computed from vectors obtained via transform.
Updated cluster to use the centralized incompatibility detection and backend switch so clustering falls back cleanly to the non-graph routine when ANN/graph is unavailable.

Testing

Compiled the modified module with python -m py_compile backend/services/intelligence/txtai_service.py, and compilation succeeded.
No additional automated tests were available in this change; runtime behavior should be validated in an environment with txtai/faiss indexes to confirm the fallback behavior prevents the previous repeat errors.

Codex Task

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/AJaySi/ALwrity/pull/403 **Author:** [@AJaySi](https://github.com/AJaySi) **Created:** 3/9/2026 **Status:** ✅ Merged **Merged:** 3/9/2026 **Merged by:** [@AJaySi](https://github.com/AJaySi) **Base:** `main` ← **Head:** `codex/fix-nprobe-attribute-error-in-search` --- ### 📝 Commits (2) - [`cd8582e`](https://github.com/AJaySi/ALwrity/commit/cd8582eb8c8f57b6d9d906403180085ae9cd0850) Fix txtai nprobe fallback to avoid reloading incompatible faiss index - [`952824a`](https://github.com/AJaySi/ALwrity/commit/952824a271529b4545a90684c084e504fb2407f6) Stabilize txtai nprobe handling without dropping loaded index state ### 📊 Changes **1 file changed** (+70 additions, -47 deletions) <details> <summary>View changed files</summary> 📝 `backend/services/intelligence/txtai_service.py` (+70 -47) </details> ### 📄 Description ### Motivation - Backend logs showed repeated `'IndexIDMap' object has no attribute 'nprobe'` errors when the service attempted to fallback from FAISS to numpy but then reloaded a persisted FAISS ANN index, causing the same failure to recur. - The goal is to make `TxtaiIntelligenceService` more resilient by ensuring backend switches do not reintroduce incompatible persisted ANN state and by providing safe retry/fallback code paths for search, similarity, and clustering. ### Description - Changed `_initialize_embeddings` to accept `load_existing_index: bool` and added logic to skip loading a persisted index when instructed (prevents reloading incompatible FAISS state). - Added `_is_nprobe_incompatibility` to centralize detection of the FAISS `IndexIDMap`/`nprobe` error pattern and `_switch_to_numpy_backend` to encapsulate backend switching and reinitialization behavior. - Hardened `search` to retry against a numpy backend with `load_existing_index=False` and, if a retry still raises the same error, force a non-ANN retrieval path by calling `search(..., index=False)`. - Hardened `get_similarity` to switch to numpy on the incompatibility and, if `similarity()` keeps failing, fall back to a cosine similarity computed from vectors obtained via `transform`. - Updated `cluster` to use the centralized incompatibility detection and backend switch so clustering falls back cleanly to the non-graph routine when ANN/graph is unavailable. ### Testing - Compiled the modified module with `python -m py_compile backend/services/intelligence/txtai_service.py`, and compilation succeeded. - No additional automated tests were available in this change; runtime behavior should be validated in an environment with txtai/faiss indexes to confirm the fallback behavior prevents the previous repeat errors. ------ [Codex Task](https://chatgpt.com/codex/tasks/task_e_69ae6a6c16c883289c58abe3ef05712b) --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

kerem

2026-03-13 21:05:04 +03:00

closed this issue
added the
pull-request
label

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

starred/ALwrity#707

No description provided.

Rows
Columns