Does the CPU architecture not support amd64?

[xfaker1]

2026-04-09 17:46:33,456 - sentence_transformers.SentenceTransformer - INFO - Load pretrained SentenceTransformer: /home/qzx/.sirchmunk/.cache/models/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
Loading weights: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 199/199 [00:00<00:00, 10702.09it/s]
BertModel LOAD REPORT from: /home/qzx/.sirchmunk/.cache/models/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
Key | Status | Details
------------------------+------------+--------
embeddings.position_ids | UNEXPECTED |

Notes:

• UNEXPECTED: can be ignored when loading from different task/architecture; not ok if you expect identical arch.

[wangxingjun778]

This warning is harmless and stems from minor architecture differences in position_ids handling; it can be safely ignored if embeddings are generated correctly.

[xfaker1]

Basically, there were no answers; it was just a repetition of my question.

[wangxingjun778]

The embedding model only affects the reuse cluster and does not impact the main search pipeline. Please provide the following configuration details:

1. LLM Configuration: Specify the exact model being used and how the service is deployed (e.g., via API or using an inference framework like vLLM).
2. Search Paths: Within the .../docs/ path, are there any contents relevant to the query?

Additionally, you can disable the LLM's thinking process to improve efficiency.

[xfaker1]

env:

# ===== Sirchmunk Configuration =====
# Generated by: sirchmunk init
# Shared by Web API server and MCP server.

# ===== LLM Settings =====
# LLM API base URL (OpenAI-compatible endpoint)
LLM_BASE_URL=https://api.minimaxi.io/v1

# LLM API key (REQUIRED - get from your LLM provider)
LLM_API_KEY=....

# LLM model name
LLM_MODEL_NAME=MiniMax-M2.7

# LLM request timeout in seconds
LLM_TIMEOUT=60.0

# ===== UI Settings =====
# WebUI theme (light/dark)
UI_THEME=light

# WebUI language (en/zh)
UI_LANGUAGE=en

# ===== Sirchmunk Settings =====
# Working directory for data and cache
SIRCHMUNK_WORK_PATH=~/.sirchmunk

# Enable verbose logging (true/false)
SIRCHMUNK_VERBOSE=false

# ===== Search Settings =====
# Default search paths (comma-separated), e.g. /data/docs,/data/pdfs
# When set, acts as the default for search(..., paths=...) if none are provided.
SIRCHMUNK_SEARCH_PATHS=/home/qzx/.sirchmunk/docs

# Maximum directory depth to search
DEFAULT_MAX_DEPTH=5

# Number of top files to return
DEFAULT_TOP_K_FILES=3

# Number of keyword granularity levels (1-5)
DEFAULT_KEYWORD_LEVELS=3

# Grep operation timeout in seconds
GREP_TIMEOUT=60.0

# ===== Cluster Settings =====
# Enable knowledge cluster reuse with embeddings
SIRCHMUNK_ENABLE_CLUSTER_REUSE=true

# Similarity threshold for cluster reuse (0.0-1.0)
CLUSTER_SIM_THRESHOLD=0.85

# Number of similar clusters to retrieve
CLUSTER_SIM_TOP_K=3

# Maximum queries per cluster (FIFO)
MAX_QUERIES_PER_CLUSTER=5

# ===== MCP Server Settings =====
# MCP server name
MCP_SERVER_NAME=sirchmunk

# Logging level (DEBUG, INFO, WARNING, ERROR, CRITICAL)
MCP_LOG_LEVEL=INFO

# MCP transport protocol (stdio or http)
MCP_TRANSPORT=stdio

# ===== HTTP Transport Settings (only used when MCP_TRANSPORT=http) =====
# Host for HTTP transport
MCP_HOST=localhost

# Port for HTTP transport
MCP_PORT=8080

# ===== Optional: Custom Embedding Model Settings =====
# Embedding model ID (from ModelScope or HuggingFace)
EMBEDDING_MODEL_ID=sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

# Embedding model cache directory
EMBEDDING_CACHE_DIR=${SIRCHMUNK_WORK_PATH}/.cache/models

The answer I need is in the "docs," but it's in Chinese. Even when I search in Chinese, the results are similar.

[wangxingjun778]

We identified the issue as a compatibility problem with the Minimax M2.7 model. In the search pipeline, steps requiring LLM participation did not yield the expected results. We will promptly fix and adapt to this issue in version 0.0.7.