Files
music_library/backlog/tasks/ml-156 - Explore-alternatives-to-reduce-token-usage-when-providing-collection-context-to-LLM-for-collection-chat.md
T
2026-05-04 21:22:27 +01:00

23 KiB

id, title, status, assignee, created_date, updated_date, labels, dependencies, references, documentation, priority
id title status assignee created_date updated_date labels dependencies references documentation priority
ML-156 Explore alternatives to reduce token usage when providing collection context to LLM for collection chat To Do
2026-05-02 16:02 2026-05-04 08:19
ready
backlog://document/doc-1
lib/music_library/chats/collection_chat.ex
lib/music_library/collection.ex
lib/music_library/chats/prompt.ex
lib/music_library/chats/stream_provider.ex
lib/open_ai/api.ex
lib/music_library_web/components/chat.ex
lib/music_library_web/live/collection_live/index.ex
high

Description

Currently, every new collection chat sends the ENTIRE collection catalog (all records formatted as "Artist - Title (year, format) [genres]") plus aggregated stats as the instructions parameter to the OpenAI Responses API. For a collection of 500+ records, this burns ~9,000+ input tokens on EVERY new chat start — regardless of what the user asks.

The goal of this task is to analyze alternatives, pick the best one, and implement it. The selected approach should:

  • Significantly reduce per-chat token usage
  • Preserve or improve the quality of LLM responses about the collection
  • Not require architectural overhauls beyond the chat/streaming layer

The Collection.collection_summary/0 function loads ALL records from the DB, formats them, and returns {summary, count}. This is computed asynchronously in CollectionLive.Index.mount/3 and passed to the Chat component as chat_context. CollectionChat.build_instructions/2 then embeds the full summary into the instructions string sent to OpenAI.

Acceptance Criteria

  • #1 CollectionChat.build_instructions/2 no longer interpolates the full collection catalog into the instructions sent to OpenAI — only aggregated stats and a record count are included
  • #2 A file_search tool with the collection's vector store ID is included in every collection chat request to the OpenAI Responses API
  • #3 A CollectionChat.FileStore module manages the collection file lifecycle: upload to OpenAI Files API, vector store creation, and file-to-store attachment, persisting IDs via Secrets
  • #4 When a record is added, edited, or deleted, the collection file at OpenAI is refreshed (async, non-blocking) so the LLM always searches up-to-date data
  • #5 If the file/vector store is unavailable (upload failed, first deploy), the chat falls back to stats-only instructions without errors
  • #6 Existing record and artist chats continue to work without changes (no regression in streaming)
  • #7 Tests cover: Files API endpoints, file upload/create/refresh lifecycle, file_search tool inclusion in chat requests, fallback when file store is unavailable, empty collection edge case
  • #8 Per-chat input tokens for the collection chat instructions are O(1) relative to collection size (the file is searched by OpenAI on demand, not embedded in instructions)

Implementation Plan

Approach: OpenAI file_search tool + Oban-managed file lifecycle

Upload the collection catalog as a file to OpenAI, create a vector store, and use the Responses API's built-in file_search tool. OpenAI automatically performs semantic search over the file and includes relevant results inline in the response stream — no SSE event handling changes, no orchestration loop.

Token savings: ~9,000 → ~100 tokens per chat (99% reduction). File search results consume ~200-500 tokens only when the model actually searches.

Research and alternative analysis: see ML-156 Research document.

Architecture decisions

  1. Chat sessions never upload or refresh files. They read vector_store_id from Secrets. If missing, they fall back to stats-only instructions (no errors surfaced to the user).
  2. File upload and refresh happen exclusively in Oban workers. No Task.start, no start_async — all OpenAI API calls for file management are serialized through a unique Oban worker with a 5-minute debounce window.
  3. Refresh is triggered at the lowest level. Records.create_record/1, Records.update_record/2, and Records.delete_record/1 enqueue the refresh worker on success. This catches all mutation paths (Index, Show, cart import, barcode scan, batch operations) without PubSub or LiveView-specific hooks.
  4. The full collection_summary/0 (catalog + stats) is eliminated from the LiveView page load. CollectionLive.Index.mount/3 loads only a lightweight stats_summary/0 — the same aggregated stats (record count, artist count, genre/format/era breakdowns) but without the per-record catalog lines. The full catalog is only loaded inside the Oban worker when a file upload/refresh is needed. Stats are ~200 tokens, always included in instructions, giving the LLM high-level collection awareness even before file search.

Phase 1: OpenAI API extensions (7 new endpoints)

Add to OpenAI.API (following existing new_request/1 pattern with Req.RateLimiter on :open_ai bucket):

  • upload_file(file_content, filename, config)POST /v1/files with multipart body (purpose: "assistants", file field containing the catalog text). Uses Req multipart support.
  • create_vector_store(name, config)POST /v1/vector_stores with %{name: name} body.
  • add_file_to_vector_store(store_id, file_id, config)POST /v1/vector_stores/{store_id}/files with %{file_id: file_id} body.
  • remove_file_from_vector_store(store_id, file_id, config)DELETE /v1/vector_stores/{store_id}/files/{file_id}. Detaches the old file from the vector store before deleting.
  • delete_file(file_id, config)DELETE /v1/files/{file_id}. Called after detaching from vector store.
  • get_file(file_id, config)GET /v1/files/{file_id}. Returns file metadata including status field (uploaded, processed, error). Used to poll indexing completion after attach.
  • list_files(config)GET /v1/files. Returns paginated list of files. Used by the periodic cleanup worker to find orphaned resources.

~90 lines.


Phase 2: FileStore module

Create MusicLibrary.Chats.CollectionChat.FileStore:

defmodule MusicLibrary.Chats.CollectionChat.FileStore do
  @moduledoc """
  Manages the collection catalog file lifecycle at OpenAI.
  Persists file_id and vector_store_id via Secrets.

  All OpenAI API calls happen inside Oban workers — never from chat
  sessions or LiveViews. The chat session path only calls get_vector_store_id/0.
  """

  @spec upload_or_refresh() :: :ok | {:error, term()}
  def upload_or_refresh do
    # 1. Check Secrets for existing file_id + vector_store_id
    # 2. If missing → initial upload flow:
    #    a. Call Collection.collection_summary/0 to get catalog text
    #    b. POST /v1/files → file_id
    #    c. POST /v1/vector_stores → vector_store_id
    #    d. POST /v1/vector_stores/{id}/files → attach
    #    e. Poll GET /v1/files/{file_id} until status == "processed"
    #       (max 10 attempts, 2s backoff)
    #    f. Persist both IDs via Secrets.store/2
    # 3. If present → refresh flow:
    #    a. Call Collection.collection_summary/0
    #    b. POST /v1/files → new_file_id
    #    c. POST /v1/vector_stores/{store_id}/files → attach new file
    #    d. Poll GET /v1/files/{new_file_id} until status == "processed"
    #    e. DELETE /v1/vector_stores/{store_id}/files/{old_file_id} → detach old
    #    f. DELETE /v1/files/{old_file_id} → delete old file
    #    g. Update file_id in Secrets (vector_store_id unchanged)
    # 4. Log errors at each step; return :ok or {:error, reason}
  end

  @spec get_vector_store_id() :: {:ok, String.t()} | {:error, :not_uploaded}
  def get_vector_store_id do
    # Read vector_store_id from Secrets
    # Return {:ok, id} or {:error, :not_uploaded}
  end

  @spec cleanup_orphaned_files() :: :ok
  def cleanup_orphaned_files do
    # Called by periodic cleanup worker
    # 1. Read known file_id + vector_store_id from Secrets
    # 2. Call GET /v1/files to list all files
    # 3. Delete any file whose id doesn't match the known file_id
    #    (with safety check: also detach from any vector store first)
    # 4. Log orphan count, return :ok
  end
end

Key design properties:

  • upload_or_refresh/0 is the single entry point called by the Oban worker — never by chat sessions
  • get_vector_store_id/0 is the single entry point called by chat sessions — read-only, no side effects
  • cleanup_orphaned_files/0 is a safety net for edge cases (e.g., Secrets write failed after OpenAI upload succeeded)

~70 lines.


Phase 3: Oban workers

3a. RefreshCollectionFile worker

Create MusicLibrary.Worker.RefreshCollectionFile:

defmodule MusicLibrary.Worker.RefreshCollectionFile do
  use Oban.Worker,
    queue: :default,
    unique: [period: 300, keys: [:worker]],
    max_attempts: 3

  @impl true
  def perform(_job) do
    case MusicLibrary.Chats.CollectionChat.FileStore.upload_or_refresh() do
      :ok -> :ok
      {:error, reason} -> {:error, reason}
    end
  end
end
  • unique: [period: 300, keys: [:worker]] — 5-minute debounce window. Only one refresh can be executing or scheduled at a time. Rapid record mutations (cart import of 50 records) each call Oban.insert but only the first insert succeeds; subsequent inserts within the 300s unique window are silently ignored by Oban.
  • max_attempts: 3 — transient OpenAI failures get 2 retries.
  • Queue :default (concurrency 10) — the worker is fast (a few API calls), no need for a dedicated queue.

Enqueued from two places:

  1. Record mutationsRecords.create_record/1, update_record/2, delete_record/1 (see Phase 4)
  2. Cron — every hour in both dev and prod, to handle initial upload after deploy and recovery after failures:
# config/prod.exs (add to crontab)
{"0 * * * *", MusicLibrary.Worker.RefreshCollectionFile}

# config/dev.exs (add to crontab)
{"*/15 * * * *", MusicLibrary.Worker.RefreshCollectionFile}

~15 lines.

3b. CleanupOrphanedOpenAIFiles worker

Create MusicLibrary.Worker.CleanupOrphanedOpenAIFiles:

  • Cron: monthly (1st of month, 11 AM)
  • Calls FileStore.cleanup_orphaned_files/0
  • Queue: :default
  • max_attempts: 2

~12 lines.


Phase 4: Records context + Collection context + LiveView changes

4a. Records context — enqueue refresh on mutations

In lib/music_library/records.ex, after each successful mutation, enqueue the refresh worker outside the with chain (non-blocking — the mutation succeeds regardless of enqueue result):

def create_record(attrs \\ %{}) do
  with {:ok, record} <- do_create_record(attrs),
       record = Enrichment.best_effort_extract_colors(record),
       :ok <- refresh_artist_info_async(record) do
    enqueue_collection_file_refresh()
    {:ok, record}
  end
end

def update_record(%Record{} = record, attrs) do
  with {:ok, updated_record} <- do_update_record(record, attrs),
       :ok <- refresh_artist_info_async(updated_record) do
    enqueue_collection_file_refresh()
    {:ok, updated_record}
  end
end

def delete_record(%Record{} = record) do
  with {:ok, record} <- Repo.delete(record) do
    record
    |> Record.artist_ids()
    |> Enum.each(fn artist_id ->
      Artists.prune_artist_info_async(artist_id)
    end)

    enqueue_collection_file_refresh()
    {:ok, record}
  end
end

defp enqueue_collection_file_refresh do
  %{} |> MusicLibrary.Worker.RefreshCollectionFile.new() |> Oban.insert()
end
  • Oban.insert/1 (not insert!) — failures to enqueue are silently ignored; the hourly cron acts as a safety net.
  • The enqueue happens after the with succeeds, ensuring we only refresh when data actually changed.
  • All mutation paths are covered: single create (AddRecord, barcode scan), cart import (batch creates via ImportFromMusicbrainzReleaseGroup worker, which calls create_record/1 internally), edit (Index and Show), delete.

~15 lines.

4b. Collection context — add stats_summary/0 and count_records/0

Add to lib/music_library/collection.ex:

@spec count_records() :: non_neg_integer()
def count_records do
  from(r in Record, where: not is_nil(r.purchased_at))
  |> Repo.aggregate(:count)
end

@spec stats_summary() :: {String.t(), non_neg_integer()}
def stats_summary do
  # Same record load as collection_summary/0 (the heavy SQL query),
  # but only computes the stats header — NO per-record catalog formatting.
  # Returns {stats_string, group_count}.
  # Example stats_string:
  #   "# Stats: 500 releases, 200 artists\nGenres: rock 150, ...\nFormats: ...\nEras: ..."
  records = from(r in Record, where: not is_nil(r.purchased_at), select: ^essential_fields()) |> Repo.all()
  groups = records |> Enum.group_by(& &1.musicbrainz_id)
  stats = build_stats(records, length(groups))
  {stats, length(groups)}
end

~10 lines. collection_summary/0 remains in the module — it's now only called by FileStore.upload_or_refresh/0 (inside the Oban worker) to generate the catalog text for file upload. stats_summary/0 extracts the lightweight stats portion for the LiveView to load (still a full table scan, but skips the expensive per-record string formatting — ~200 tokens of output vs ~9,000).

4c. CollectionLive.Index — load stats instead of full catalog

Replace the heavy start_async(:collection_summary, &Collection.collection_summary/0) with stats_summary:

# In mount/3:
|> assign(:collection_stats, {"", 0})
|> start_async(:collection_stats, fn -> Collection.stats_summary() end)

# Rename handle_async:
def handle_async(:collection_stats, {:ok, stats}, socket) do
  {:noreply, assign(socket, :collection_stats, stats)}
end

def handle_async(:collection_stats, {:exit, _reason}, socket) do
  {:noreply, socket}
end

# In the template — pass stats as chat_context:
chat_context={@collection_stats}

The @collection_summary assign is removed. The Chat component receives {stats_string, record_count}, matching the existing tuple shape so no Chat component changes are needed.

~8 lines (mostly renames).


Phase 5: Chat streaming changes

5a. OpenAI facade — accept :vector_store_ids option

Modify OpenAI.chat_stream/2 to pass tools to the API layer:

def chat_stream(messages, opts) when is_list(messages) do
  model = Keyword.get(opts, :model, "gpt-4.1")
  temperature = Keyword.get(opts, :temperature, 0.7)
  instructions = Keyword.get(opts, :instructions, "")
  on_chunk = Keyword.fetch!(opts, :on_chunk)
  vector_store_ids = Keyword.get(opts, :vector_store_ids, [])

  tools = build_tools(vector_store_ids)

  API.chat_stream(messages, instructions, model, temperature, open_ai_config(), on_chunk, tools)
end

defp build_tools([]), do: [%{type: "web_search_preview"}]
defp build_tools(ids) when is_list(ids) do
  [%{type: "file_search", vector_store_ids: ids}, %{type: "web_search_preview"}]
end

~10 lines.

5b. OpenAI.API — accept optional tools parameter

Add a 7-arity overload to API.chat_stream:

def chat_stream(messages, instructions, model, temperature, config, cb) do
  chat_stream(messages, instructions, model, temperature, config, cb, [%{type: "web_search_preview"}])
end

def chat_stream(messages, instructions, model, temperature, config, cb, tools) when is_list(tools) do
  config
  |> new_request()
  |> Req.merge(
    url: "/v1/responses",
    receive_timeout: 60_000,
    connect_options: [timeout: 5_000],
    json: %{
      model: model,
      instructions: instructions,
      input: messages,
      tools: tools,
      stream: true,
      temperature: temperature
    },
    into: :self
  )
  |> do_chat_stream(cb)
end
  • Uses function clause pattern matching (when is_list(tools)) rather than default argument with ||, avoiding the falsy-empty-list pitfall.
  • The 6-arity version is preserved for backward compatibility (existing RecordChat and ArtistChat callers).

~5 lines.

5c. CollectionChat — new instructions and streaming logic

Update CollectionChat:

@impl true
def stream_response(messages, {stats, record_count}, callback) do
  instructions = build_instructions(stats, record_count)

  vector_store_opts =
    case FileStore.get_vector_store_id() do
      {:ok, id} -> [vector_store_ids: [id]]
      {:error, :not_uploaded} -> []  # fall back to stats-only
    end

  OpenAI.chat_stream(messages, [
    on_chunk: callback,
    instructions: instructions,
    model: "gpt-5.1"
  ] ++ vector_store_opts)
end

defp build_instructions(stats, record_count) do
  Prompt.build("""
  Answer questions about the user's music collection.
  Use the provided stats and file search to answer questions
  about the collection.

  #{stats}

  The collection contains #{record_count} records in total.

  Use file search to find specific records when the user asks about
  artists, albums, genres, or formats in their collection.

  If file search returns no results for a query, say so honestly —
  don't guess or invent records. If you're unsure whether a record
  exists, mention that it wasn't found in the collection.

  # Mentioning artists/albums

  **IF YOU MENTION AN ARTIST NAME OR ALBUM NAME, wrap it in "[[name]]",
  for example "[[Steven Wilson]]"
  """)
end

Key changes from current code:

  • stream_response/3 keeps the {tuple, count} signature shape — the stats string replaces the catalog text
  • build_instructions/2 interpolates aggregated stats (~200 tokens) instead of the full catalog (~9,000 tokens)
  • FileStore.get_vector_store_id/0 is the only FileStore function called from the chat path
  • On {:error, :not_uploaded}, vector_store_opts is empty → build_tools([]) returns only web_search_preview → stats-only fallback (still useful for aggregate questions)
  • Added explicit guidance for when file search returns no results
  • The "Collection catalog:" section is removed; replaced with stats inline

~18 lines (modifications to existing code).


Phase 6: Testing

6a. OpenAI.API endpoint tests (test/open_ai/api_test.exs)

Add test blocks for the 7 new endpoints, following existing Req.Test.stub patterns:

  • upload_file/3 — verify multipart body includes purpose: "assistants" and file content; test success (200 + file JSON) and error (500) responses
  • create_vector_store/2 — verify request body has name field; test success and error
  • add_file_to_vector_store/3 — verify URL path includes store_id; test success and error
  • remove_file_from_vector_store/3 — verify DELETE request; test 200 and 404 (already removed)
  • delete_file/2 — verify DELETE request; test success
  • get_file/2 — test status: "processed" (success), status: "uploaded" (still processing), and error
  • list_files/1 — test paginated response with file list
  • Rate-limit error handling (429 → ErrorResponse with :rate_limit kind) for each endpoint

~80 lines.

6b. FileStore tests (test/music_library/chats/collection_chat/file_store_test.exs)

Test FileStore functions with Req.Test stubs and direct Secrets manipulation:

  • upload_or_refresh/0 — first-time upload (no secrets → creates file + store + attaches → secrets populated)
  • upload_or_refresh/0 — refresh (secrets present → uploads new file → attaches → detaches old → deletes old → updates file_id in secrets)
  • upload_or_refresh/0 — handles file still "uploaded" (polls until "processed")
  • upload_or_refresh/0 — handles OpenAI API errors gracefully (returns {:error, ...})
  • get_vector_store_id/0 — returns {:ok, id} when secret exists
  • get_vector_store_id/0 — returns {:error, :not_uploaded} when secret missing
  • cleanup_orphaned_files/0 — deletes files not matching known file_id; skips known file

~70 lines.

6c. CollectionChat tests (test/music_library/chats/collection_chat_test.exs)

Update existing tests and add new ones:

  • Verify instructions no longer contain the per-record catalog lines (the key regression test)
  • Verify instructions do contain the aggregated stats string and record count
  • Verify file_search tool is included in the API request body when vector store is available (stub FileStore.get_vector_store_id/0{:ok, "vs_test"} and assert tools array in the JSON body contains %{"type" => "file_search", "vector_store_ids" => ["vs_test"]})
  • Verify file_search tool is not included when vector store is unavailable (stub → {:error, :not_uploaded}, assert tools only has web_search_preview)
  • Verify web_search_preview is always included regardless of file_search presence
  • Test empty collection edge case (stats = "", record_count = 0)
  • Test SSE event handling: include response.file_search_call.in_progress, response.file_search_call.searching, and response.file_search_call.completed events in the test stream fixture to verify they're silently consumed (caught by the "response." <> _ catch-all) without breaking the stream

~60 lines (modifications to existing + new tests).

6d. Worker tests (test/music_library/worker/refresh_collection_file_test.exs)

  • perform/1 delegates to FileStore.upload_or_refresh/0 and returns :ok
  • perform/1 returns {:error, reason} on failure
  • Unique constraint: inserting two jobs within 300s results in only one executing (test with Oban.insert/1 returning {:ok, _} for first and {:ok, %{conflict?: true}} for second)

~20 lines.

6e. Records context integration tests

In existing test/music_library/records_test.exs (or equivalent):

  • create_record/1 enqueues RefreshCollectionFile worker
  • update_record/2 enqueues RefreshCollectionFile worker
  • delete_record/1 enqueues RefreshCollectionFile worker
  • Verify with assert_enqueued (Oban testing helper)

~15 lines.

6f. LiveView test

In test/music_library_web/live/collection_live/index_test.exs (or equivalent):

  • Verify @collection_stats is set after mount (not @collection_summary)
  • Verify Chat component receives {stats_string, count} tuple as chat_context

~8 lines (modifications to existing test).

6g. Collection context tests

In existing test/music_library/collection_test.exs (or equivalent):

  • stats_summary/0 returns stats string with genre/format/era breakdowns and correct group count
  • stats_summary/0 returns {"", 0} for empty collection
  • count_records/0 returns correct count
  • Existing collection_summary/0 tests continue to pass unchanged

~20 lines.

Estimated total: ~510 lines across 13-15 files. Risk: low/medium — the core streaming infrastructure (decode_responses_event/2, Chat component do_send_message/2, SSE event loop) is untouched. New code is additive (API endpoints, Oban worker, FileStore module) and follows existing patterns.

Definition of Done

  • #1 All new and modified modules have @moduledoc
  • #2 All public functions have @spec and @doc
  • #3 Mix compile --warnings-as-errors passes
  • #4 mix test passes with no failures
  • #5 mix format --check-formatted passes
  • #6 mix credo passes
  • #7 Documentation updated (architecture.md if new modules/schemas added)
  • #8 Commit subject references ML-156