My PAWDOC work archive of some 17,000 Index entries and 31,000 associated files has an Index in the FileMaker database app. That was the app I chose in 1986 when I moved the index from a card-based system \u2013 and I\u2019ve stuck with FileMaker ever since despite it becoming a hugely more complex and capable beast. A few months ago, I discovered that FileMaker was offering a free 45-day trial of its latest AI-enabled version, so I thought I\u2019d try it out as part of this investigation into AI for personal archives. I figured that it would help me understand a bit more about AI, and that perhaps what FileMaker was doing would be generalisable to all specialist archive-supporting apps. This post summarises what I discovered and concluded.<\/p>\n
There is an extensive range of videos and training materials<\/a> for FileMaker available online. I had already looked at a number of the videos<\/a> as part of my Phase 4 investigations, and now I went through two of the training modules:<\/p>\n These describe the following three AI-related features:<\/p>\n These features have been built into FileMaker using about a dozen AI-related FileMaker Scripts – sets of predefined instructions carried out by the application when a particular script is run by the user. \u00a0The Semantic Search<\/strong> works by first generating the vector value of the contents of a field (see Terminology in the post at this link<\/a>), putting that value into a file, and storing that file with the relevant record in the database. This is known as embedding and is enabled by one of the pre-built AI scripts called \u2018Insert Embedding in Found Set\u2019. Another script (Perform Semantic Find) turns a natural language search term input by the user into a specific vector value, and compares that vector value with the vector values for the records which have already been given vector values. The result of the comparison is a Cosine Similarity Coefficient ranging between +1 and -1 and typically calculated to 6 or 7 decimal points. FileMaker then lists the top X records which have the highest Cosine Similarity score (where X is a user-specified variable). A more sophisticated version of this is a hybrid version of this search which provides a screen in which the user defines BOTH a search term for the FileMaker Find facility to derive a subset, AND a natural language search request to be used on the contents of a specific field in all the records in the subset. In both the Ordinary Semantic Search and the Hybrid Semantic Search, the vector embedding files must already be in place in the database: records without a vector embedding file will not be included in this search. As with all AI usage, there may be limits or cost implications related to the number\/size of requests that are made to the model.<\/p>\n The Generate Response from Model<\/strong> feature enables a general-purpose interface to be created with a local or remote AI model. This can be used to gather information from the database and for that information to be sent with an instruction Prompt to the model. The response from the AI is then displayed in FileMaker in a pre-specified way. For example, information in certain fields in a Product Database could be sent to the AI together with instructions to provide single line summaries of each product consisting of short description, category and price. The specific information to be gathered, and the instructions, could all be pre-specified in particular scripts, or could be defined by the user. Inevitably, the scripts to do this may be a little complicated; and appropriate interface screens (\u2018layouts\u2019 in FileMaker terms) will need to be built. As with the Semantic Search feature, there may be limits or cost implications related to the number\/size of requests that are made to the model.<\/p>\n Unlike the previous two features, the Perform Find by Natural Language <\/strong>facility does not send any of the database contents to an AI model. Instead, it sends the prompt request created by the user along with further instructions plus the database schema. The AI then creates and returns the Find text which FileMaker includes in the relevant fields before running a Find operation.<\/p>\n The training courses did a good job of explaining how all three of these features work. I then tried out the Semantic Find function on my PAWDOC database using the trial 2026 version of FileMaker. This involved creating additional scripts<\/a> around the primary scripts to specify things like the AI model being used, the fields for storing inputs and outputs, and the particular fields to be interrogated in the database. I initially followed the example of the training courses and obtained a trial API key to be able to access models in the cloud from the Cohere provider. However, I also followed the steps described in a very informative blog post called ‘Using LM Studio as a local LLM for FileMaker AI Development<\/a>‘ which showed how the same three features could be used with a local model residing on your own local machine. The examples in the blog post use the LM Studio app which I happened to already have on my laptop. So, I ended up carrying out a set of tests using both a remote and a local AI model. The tests were performed on the PAWDOC Title field to a) embed the contents, and b) perform semantic searches on a subset of the PAWDOC records. The models used were LM Studio\u2019s “text-embedding-nomic-embed-text-v2-moe” on my laptop, and Cohere\u2019s “embed-english-light-v3.0” in the cloud. \u00a0For comparison purposes, I also conducted some equivalent searches using the standard FileMaker Find function. My observations from these tests<\/a> include the following:<\/p>\n Embedding<\/em><\/p>\n Quality of Search Results<\/em><\/p>\n Comparison of the hits produced by LM Studio and Cohere<\/em><\/p>\n Cosine Similarity Scores<\/em><\/p>\n Comparison with Standard FileMaker Find Results<\/em><\/p>\n While the tests from which the above observations are drawn were by no means rigorous or definitive, they do provide enough insights to draw the following general conclusions:<\/p>\n Below is a breakdown of the time I spent on Phase 6 and across all phases.<\/p>\n My PAWDOC work archive of some 17,000 Index entries and 31,000 associated files has an Index in the FileMaker database app. That was the app I chose in 1986 when I moved the index from a card-based system \u2013 and … Continue reading \n
\n
\n
\n
\n
\n
\n
\n
\n\n
\n Activity<\/u><\/em><\/td>\n No of Tasks or task breakdown<\/u><\/em><\/td>\n Elapsed time<\/u><\/em><\/td>\n Time spent<\/u><\/em><\/td>\n<\/tr>\n \n Phase 1<\/td>\n 70 (started 05Mar2026)<\/u><\/td>\n 43 days<\/td>\n 105 hrs<\/td>\n<\/tr>\n \n Phase 2<\/td>\n 8<\/td>\n 4 days<\/td>\n 11 hrs<\/td>\n<\/tr>\n \n Phase 3<\/td>\n \u00b7 Create test files, test, analyse results
\n\u00b7 Research & draft pwofc.com posts<\/td>\n3 days
\n4 days<\/td>\n15 hrs
\n12 hrs<\/td>\n<\/tr>\n\n Phase 4<\/td>\n \u00b7 Create test files, test, analyse results
\n\u00b7 Research & draft pwofc.com posts<\/td>\n14 days
\n11 days<\/td>\n80 hrs
\n31 hrs<\/td>\n<\/tr>\n\n Phase 5<\/td>\n \u00b7 Create test files, test, analyse results
\n\u00b7 Research & draft pwofc.com posts<\/td>\n4 days
\n2 days<\/td>\n13 hrs
\n8 hrs<\/td>\n<\/tr>\n\n Phase 6<\/td>\n \u00b7 Do training & draft pwofc.com post
\n\u00b7 Create test files, test, analyse results<\/td>\n7 days
\n6 days<\/td>\n16 hrs
\n12 hrs<\/td>\n<\/tr>\n\n Totals<\/td>\n <\/td>\n 85 days<\/td>\n \u00a0303 hrs<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n","protected":false},"excerpt":{"rendered":"