What should I do to make the program support the model I want to replace?

[TaliaRosenberg]

This open source project is truly amazing. Thank you for your work. However, when I try to replace some models—for example, when I swap BAAI/bge-small-en-v1.5 with BAAI/bge-large-en-v1.5—the system throws an error: "Model BAAI/bge-large-en-v1.5 produces 1024-dimensional embeddings, but database schema requires 384 dimensions. Use a model that produces 384-dimensional embeddings." What should I do to make the program support the model I want to replace?

[nicoloboschi]

Hey @TaliaRosenberg thanks for trying out Hindsight!

At the moment we require a fixed vector dimension, but we could support more dimensions if we want.

Is there a specific reason you want that embedding model?

[TaliaRosenberg]

@nicoloboschi Thanks for getting back to me even though you're swamped. We need to support languages like German, Japanese, and other Asian languages too.

[husniadil]

@nicoloboschi I am wondering how would you support more dimensions? Having multiple tables like memory_units_1024, memory units_1536, etc? How is the best approach?

I made a change on my fork to integrate hindsight with openai embedding models, it works with 384 dimensions, but it may degrade the output quality. Let me know if this is worth to PR this into hindsight main repo, otherwise I will be just for me use case: husniadil@f1821ef

I have tested it with a criminal investigation scenarios and it works great

⏺ Criminal Investigation Case Test Results

  The Hindsight API works excellently for criminal investigation scenarios. Here's a summary:

  Data Ingested

  | Category              | Items                                     |
  |-----------------------|-------------------------------------------|
  | Case Overview         | 1                                         |
  | Suspect Profiles      | 3 (Marcus Webb, Victor Chen, Diana Reyes) |
  | Witness Statements    | 3                                         |
  | Physical Evidence     | 3                                         |
  | Digital Evidence      | 2                                         |
  | Financial Evidence    | 1                                         |
  | Investigation Updates | 3                                         |
  | Alibi Verifications   | 2                                         |

  Key Capabilities Demonstrated

  Entity Recognition - Automatically extracted 20+ entities including:
  - Suspects: Marcus Webb (6 mentions), Diana Reyes (5 mentions), Victor Chen (4 mentions)
  - Locations: Main Street, Riverside Jewelry, Nicks Bar, Highway 101
  - Evidence: NightOps tactical jacket, burner phone, $12,000 cash
  - Vehicles: license plate 7XK-492, Eastside Auto Rental

  Relationship Mapping - Connected:
  - Marcus Webb ↔ Victor Chen (known associates)
  - Victor Chen ↔ Robert Chen (cousins) ↔ Eastside Auto Rental
  - Diana Reyes ↔ Riverside Jewelry (former employee)
  - Evidence linking suspects to the crime scene

  Semantic Search - Successfully answered:
  - "What evidence connects Marcus Webb to the crime?"
  - "What is the connection between Victor Chen and the getaway vehicle?"

  Case Analysis (Reflect) - Provided intelligent reasoning:
  - Identified Marcus Webb as the likely perpetrator based on evidence
  - Analyzed Diana Reyes's potential role as an insider
  - Considered alibis and their weaknesses

  Temporal Reasoning - Correctly tracked timeline:
  - Dec 14 9PM: Marcus & Diana meet at Nick's Bar
  - Dec 15 2:15AM: Van spotted behind jewelry store
  - Dec 15 2:30AM: Burglary occurs
  - Dec 15 2:45AM: Glass breaking heard, van flees
  - Dec 15 3:45AM: Van spotted on Highway 101

  The API successfully handles complex, multi-entity, temporally-linked investigation data with semantic search and AI-powered analysis.

Using the same memory bank, I can recall it using different languages.

⏺ Bash(curl -s -X POST 'http://localhost:8888/v1/default/banks/case-001/memories/recall' -H
      'Content-Type: application/json' -d '{"query": "证据显示谁是主要嫌疑人？", "budget": "low"}' | jq
      '.results[:2]')
  ⎿  PreToolUse:Callback hook succeeded: Success
  ⎿ [
      {
        "id": "49f8a753-056d-4e79-b0bf-6b7a54f81561",
        "text": "User believes that more evidence and details are needed before reaching a final
    conclusion about the involvement of certain individuals in the robbery | When: Sunday, December
     28, 2025 | Involving: user | User feels that without sufficient evidence, it is premature to
    make a judgment about the individuals' involvement in the robbery, indicating a cautious and
    analytical approach to the situation",
        "type": "opinion",
        "entities": [
          "user",
          "robbery",
          "evidence",
    "details"
        ],
        "context": "formed during thinking about:
    この宝石強盗事件の主犯は誰だと思いますか？理由も教えてください。",
        "occurred_start": null,
        "occurred_end": null,
        "mentioned_at": "2025-12-28T05:27:15.523174+00:00",
        "document_id": "7b687d0f-17fe-4829-82ee-4269f9addff6",
        "metadata": null,
        "chunk_id": "case-001_7b687d0f-17fe-4829-82ee-4269f9addff6_0"
      },
      {
        "id": "da54b8e3-f840-4493-9f95-b55ffb98e3a2",
        "text": "User thinks Marcus is the most suspicious individual in the jewel robbery case due
     to the lack of direct evidence against Diana Reyes, despite indications of her involvement. |
    When: Sunday, December 28, 2025 | Involving: User, Marcus (suspect in the jewel robbery case),
    Diana Reyes (indicated to be involved in the case) | User believes that while there are
    indications of Diana Reyes's involvement, the absence of direct evidence makes it unjust to
    solely blame her, leading to suspicion towards Marcus instead.",
        "type": "opinion",
        "entities": [
          "user",
          "Marcus",
          "Diana Reyes",
    "jewel robbery case"
        ],
        "context": "formed during thinking about:
    この宝石強盗事件の主犯は誰だと思いますか？理由も教えてください。",
        "occurred_start": null,
        "occurred_end": null,
        "mentioned_at": "2025-12-28T05:27:15.523174+00:00",
        "document_id": "caa572a2-1c21-4ac0-a1a0-ee1bf12a77d3",
        "metadata": null,
        "chunk_id": "case-001_caa572a2-1c21-4ac0-a1a0-ee1bf12a77d3_0"
      }
    ]

  ⎿  PostToolUse:Callback hook succeeded: Success

It can even find the suspect using /reflect, which is fantastic.

⏺ Bash(curl -s -X POST 'http://localhost:8888/v1/default/banks/case-001/reflect' -H 'Content-Type:
      application/json' -d '{"query": "Siapa pelaku perampokan Riverside Jewelry?"}' | jq -r '.text')
  ⎿  PreToolUse:Callback hook succeeded: Success
  ⎿ Based on the analysis and the evidence surrounding the Riverside Jewelry heist, I believe that
    Marcus Webb stands out as the primary suspect. My reasoning stems from several factors.

    Firstly, there’s a significant amount of circumstantial evidence that points towards Marcus,
    especially considering his likely role as the mastermind behind the heist. His relationship
    with Diana, coupled with the meetings they had before the robbery, suggests a level of
    premeditation that cannot be overlooked. This implies a deeper involvement in orchestrating the
     crime.

    Diana Reyes, while having valuable insight into the store due to her previous employment,
    presents a weaker case in terms of direct evidence against her. Although her background
    certainly raises questions, particularly regarding her previous termination for theft, the lack
     of concrete proof linking her directly to the robbery makes it difficult to place the same
    level of suspicion on her.

    It's also essential to consider the nature of their alibis and the overall context of their
    actions. Diana’s alibi has significant gaps and is not entirely credible, but the investigative
     focus seems to have gravitated more towards Marcus, especially as the evidence and
    circumstances surrounding him appear stronger and more compelling.

    In summary, while I believe that both individuals had roles in the robbery, Marcus Webb is
    likely the key suspect based on the existing evidence and the dynamics of their relationship.
    However, I emphasize the importance of ongoing investigation, as new evidence could alter this
    understanding, reinforcing my belief in approaching conclusions with caution.

[nicoloboschi]

@husniadil my current idea is to have this as a parameter and create a dedicated column for that dimension.

In case you change that parameter after you already have some data, we either need to ask the user to migrate (maybe a quick command from the cli) or reset the database

Im going to work on it in the next few days