This paper proposes a Bag of Visual Words (BoVW) based approach for keyword spotting on the Mongolian historical document images. In this paper, the first step is dividing the scanned Mongolian historical document images into word images by some preprocessing steps, such as connected component analysis, binarization etc. Then, all of image in our training set are processed in the following steps, including extracting keypoints, obtaining local descriptors and formulating visual word. Finally, each word image can be represented as a histogram of visual words by a codebook. In the retrieval stage, a provided query keyword image is also converted into a histogram of visual words through the above-mentioned procedure. After that, similarities between a query keyword image and whole candidate of word images can be calculated. Therefore, a sorted list will be returned in descending order of the similarities. Moreover, spatial information of visual word is introduced into the original framework of BoVW by the spatial pyramid matching (SPM) technology. Experimental results show that addition of spatial information obtains a good performance on our dataset.