package ocaml-ai-sdk

  1. Overview
  2. Docs
OCaml AI SDK - Provider abstraction for AI models

Install

dune-project
 Dependency

Authors

Maintainers

Sources

ocaml-ai-sdk-0.1.tbz
sha256=467ed85a42617ce399d0032d81f27f8dd2e8ca9b0753daeb2cbd84bf46761370
sha512=f5e96feea16c6ffdb45f49ee3e84e9cdb5426245c1ac32acc6434c15b949cc576c1007ed463bf1362dce507d10700eb96acf5a42fa7eed4e8e491ae0214a2e4a

doc/src/ocaml-ai-sdk.ai_core/stream_text.ml.html

Source file stream_text.ml

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
(* ID counter for stream blocks *)
type id_gen = {
  mutable text_count : int;
  mutable reasoning_count : int;
  mutable approval_count : int;
}

let make_id_gen () = { text_count = 0; reasoning_count = 0; approval_count = 0 }

let next_text_id gen =
  gen.text_count <- gen.text_count + 1;
  Printf.sprintf "txt_%d" gen.text_count

let next_reasoning_id gen =
  gen.reasoning_count <- gen.reasoning_count + 1;
  Printf.sprintf "rsn_%d" gen.reasoning_count

let next_approval_id gen =
  gen.approval_count <- gen.approval_count + 1;
  Printf.sprintf "appr_%d" gen.approval_count

(** Consume a provider stream for one step, emitting [Text_stream_part.t] events.
    Returns the accumulated text, reasoning, tool calls, finish reason, and usage. *)
let consume_provider_stream ~id_gen ~push ~on_chunk ?(on_text_accumulated = fun (_ : string) -> ()) provider_stream =
  let text_buf = Buffer.create 256 in
  let reasoning_buf = Buffer.create 256 in
  let current_text_id = ref None in
  let current_reasoning_id = ref None in
  (* Track tool call deltas for accumulation *)
  let tool_calls : (string, string * Buffer.t) Hashtbl.t = Hashtbl.create 4 in
  let completed_tool_calls = ref [] in
  let finish_reason = ref Ai_provider.Finish_reason.Unknown in
  let usage = ref { Ai_provider.Usage.input_tokens = 0; output_tokens = 0; total_tokens = None } in
  let emit part =
    push (Some part);
    match on_chunk with
    | Some f -> f part
    | None -> ()
  in
  let close_text () =
    match !current_text_id with
    | Some id ->
      emit (Text_stream_part.Text_end { id });
      current_text_id := None
    | None -> ()
  in
  let close_reasoning () =
    match !current_reasoning_id with
    | Some id ->
      emit (Text_stream_part.Reasoning_end { id });
      current_reasoning_id := None
    | None -> ()
  in
  let%lwt () =
    Lwt_stream.iter
      (fun (part : Ai_provider.Stream_part.t) ->
        match part with
        | Stream_start _ -> ()
        | Text { text } ->
          let id =
            match !current_text_id with
            | Some id -> id
            | None ->
              close_reasoning ();
              let id = next_text_id id_gen in
              emit (Text_stream_part.Text_start { id });
              current_text_id := Some id;
              id
          in
          Buffer.add_string text_buf text;
          on_text_accumulated (Buffer.contents text_buf);
          emit (Text_stream_part.Text_delta { id; text })
        | Reasoning { text } ->
          let id =
            match !current_reasoning_id with
            | Some id -> id
            | None ->
              close_text ();
              let id = next_reasoning_id id_gen in
              emit (Text_stream_part.Reasoning_start { id });
              current_reasoning_id := Some id;
              id
          in
          Buffer.add_string reasoning_buf text;
          emit (Text_stream_part.Reasoning_delta { id; text })
        | Tool_call_delta { tool_call_id; tool_name; args_text_delta; _ } ->
          close_text ();
          close_reasoning ();
          let buf =
            match Hashtbl.find_opt tool_calls tool_call_id with
            | Some (_, buf) -> buf
            | None ->
              let buf = Buffer.create 64 in
              Hashtbl.replace tool_calls tool_call_id (tool_name, buf);
              buf
          in
          Buffer.add_string buf args_text_delta;
          emit (Text_stream_part.Tool_call_delta { tool_call_id; tool_name; args_text_delta })
        | Tool_call_finish { tool_call_id } ->
          (match Hashtbl.find_opt tool_calls tool_call_id with
          | Some (tool_name, buf) ->
            let args_str = Buffer.contents buf in
            let args = Core_tool.safe_parse_json_args args_str in
            completed_tool_calls := { Generate_text_result.tool_call_id; tool_name; args } :: !completed_tool_calls;
            emit (Text_stream_part.Tool_call { tool_call_id; tool_name; args });
            Hashtbl.remove tool_calls tool_call_id
          | None -> ())
        | Finish { finish_reason = fr; usage = u } ->
          close_text ();
          close_reasoning ();
          finish_reason := fr;
          usage := u
        | Error { error } -> emit (Text_stream_part.Error { error = Ai_provider.Provider_error.to_string error })
        | File _ | Provider_metadata _ -> ())
      provider_stream
  in
  Lwt.return
    (Buffer.contents text_buf, Buffer.contents reasoning_buf, List.rev !completed_tool_calls, !finish_reason, !usage)

let stream_text ~model ?system ?prompt ?messages ?tools ?(tool_choice : Ai_provider.Tool_choice.t option)
  ?(output : (Yojson.Basic.t, Yojson.Basic.t) Output.t option) ?(max_steps = 1) ?stop_when ?max_output_tokens
  ?temperature ?top_p ?top_k ?stop_sequences ?seed ?headers ?provider_options ?on_step_finish ?on_chunk ?on_finish
  ?(pending_tool_approvals = []) () =
  (* Build initial messages *)
  let initial_messages = Prompt_builder.resolve_messages ?system ?prompt ?messages () in
  let mode = Output.mode_of_output output in
  let tools = Option.value ~default:[] tools in
  let provider_tools = Prompt_builder.tools_to_provider tools in
  (* Create output streams *)
  let full_stream, full_push = Lwt_stream.create () in
  let text_stream, text_push = Lwt_stream.create () in
  let partial_output_stream, partial_output_push = Lwt_stream.create () in
  (* Promises for final values *)
  let usage_promise, usage_resolver = Lwt.wait () in
  let finish_promise, finish_resolver = Lwt.wait () in
  let steps_promise, steps_resolver = Lwt.wait () in
  let output_promise, output_resolver = Lwt.wait () in
  (* Partial output deduplication *)
  let last_partial_json = ref "" in
  let on_text_accumulated =
    match output with
    | Some o when Option.is_some o.Output.response_format ->
      fun accumulated ->
        (match o.Output.parse_partial accumulated with
        | Some json ->
          let json_str = Yojson.Basic.to_string json in
          (match String.equal json_str !last_partial_json with
          | true -> ()
          | false ->
            last_partial_json := json_str;
            partial_output_push (Some json))
        | None -> ())
    | _ -> fun (_ : string) -> ()
  in
  let id_gen = make_id_gen () in
  (* Wrapper that also pushes text to text_stream *)
  let push_full part =
    full_push part;
    match part with
    | Some (Text_stream_part.Text_delta { text; _ }) -> text_push (Some text)
    | None -> text_push None
    | _ -> ()
  in
  (* Background streaming loop *)
  Lwt.async (fun () ->
    let emit_event part =
      push_full (Some part);
      Option.iter (fun f -> f part) on_chunk
    in
    let execute_and_emit (tc : Generate_text_result.tool_call) =
      let%lwt tr = Core_tool.execute_tool ~tools ~tool_call_id:tc.tool_call_id ~tool_name:tc.tool_name ~args:tc.args in
      emit_event
        (Text_stream_part.Tool_result
           { tool_call_id = tr.tool_call_id; tool_name = tr.tool_name; result = tr.result; is_error = tr.is_error });
      Lwt.return tr
    in
    let finish_stream ~finish_reason ~usage ~all_steps =
      emit_event (Text_stream_part.Finish { finish_reason; usage });
      push_full None;
      let parsed_output = Output.parse_output output all_steps in
      partial_output_push None;
      Lwt.wakeup_later usage_resolver usage;
      Lwt.wakeup_later finish_resolver finish_reason;
      Lwt.wakeup_later steps_resolver all_steps;
      Lwt.wakeup_later output_resolver parsed_output;
      Option.iter
        (fun f ->
          f
            {
              Generate_text_result.text = Generate_text_result.join_text all_steps;
              reasoning = Generate_text_result.join_reasoning all_steps;
              tool_calls = List.concat_map (fun (s : Generate_text_result.step) -> s.tool_calls) all_steps;
              tool_results = List.concat_map (fun (s : Generate_text_result.step) -> s.tool_results) all_steps;
              steps = all_steps;
              finish_reason;
              usage;
              response = { id = None; model = None; headers = []; body = `Null };
              warnings = [];
              output = parsed_output;
            })
        on_finish;
      Lwt.return_unit
    in
    emit_event Text_stream_part.Start;
    let rec step_loop ~current_messages ~steps ~total_usage ~step_num =
      if step_num > max_steps then
        finish_stream ~finish_reason:(Ai_provider.Finish_reason.Other "max_steps") ~usage:total_usage
          ~all_steps:(List.rev steps)
      else begin
        emit_event Text_stream_part.Start_step;
        let opts =
          Prompt_builder.make_call_options ~messages:current_messages ~tools:provider_tools ?tool_choice ~mode
            ?max_output_tokens ?temperature ?top_p ?top_k ?stop_sequences ?seed ?provider_options ?headers ()
        in
        let%lwt stream_result = Ai_provider.Language_model.stream model opts in
        let%lwt text, reasoning, tool_calls, fr, step_usage =
          consume_provider_stream ~id_gen ~push:push_full ~on_chunk ~on_text_accumulated stream_result.stream
        in
        let new_total = Generate_text_result.add_usage total_usage step_usage in
        let has_tool_calls =
          match tool_calls with
          | [] -> false
          | _ :: _ -> true
        in
        let should_continue =
          has_tool_calls
          && step_num < max_steps
          &&
          match tool_choice with
          | Some Ai_provider.Tool_choice.None_ -> false
          | Some Auto | Some Required | Some (Specific _) | None -> true
        in
        if should_continue then begin
          let%lwt blocked_calls, executable_calls = Core_tool.evaluate_approvals ~tools tool_calls in
          let%lwt tool_results = Lwt_list.map_s execute_and_emit executable_calls in
          (* Emit approval requests only for tools that have needs_approval (not client-only tools) *)
          List.iter
            (fun (tc : Generate_text_result.tool_call) ->
              match List.assoc_opt tc.tool_name tools with
              | Some { Core_tool.needs_approval = Some _; _ } ->
                let approval_id = next_approval_id id_gen in
                emit_event
                  (Text_stream_part.Tool_approval_request
                     { approval_id; tool_call_id = tc.tool_call_id; tool_name = tc.tool_name; args = tc.args })
              | _ -> ())
            blocked_calls;
          let step : Generate_text_result.step =
            { text; reasoning; tool_calls; tool_results; finish_reason = fr; usage = step_usage }
          in
          Option.iter (fun f -> f step) on_step_finish;
          emit_event (Text_stream_part.Finish_step { finish_reason = fr; usage = step_usage });
          match blocked_calls with
          | _ :: _ ->
            (* Some tools need approval — stop the stream *)
            finish_stream ~finish_reason:fr ~usage:new_total ~all_steps:(List.rev (step :: steps))
          | [] ->
            (* All tools executed — check stop conditions before continuing *)
            let%lwt stop_with_steps =
              match stop_when with
              | Some conditions ->
                let all_steps_so_far = List.rev (step :: steps) in
                let%lwt met = Stop_condition.is_met conditions ~steps:all_steps_so_far in
                Lwt.return (if met then Some all_steps_so_far else None)
              | None -> Lwt.return None
            in
            (match stop_with_steps with
            | Some all_steps_so_far -> finish_stream ~finish_reason:fr ~usage:new_total ~all_steps:all_steps_so_far
            | None ->
              let assistant_content =
                let parts = ref [] in
                if String.length text > 0 then parts := Ai_provider.Content.Text { text } :: !parts;
                List.iter
                  (fun (tc : Generate_text_result.tool_call) ->
                    parts :=
                      Ai_provider.Content.Tool_call
                        {
                          tool_call_type = "function";
                          tool_call_id = tc.tool_call_id;
                          tool_name = tc.tool_name;
                          args = Yojson.Basic.to_string tc.args;
                        }
                      :: !parts)
                  tool_calls;
                List.rev !parts
              in
              let updated_messages =
                Prompt_builder.append_assistant_and_tool_results ~messages:current_messages ~assistant_content
                  ~tool_results
              in
              step_loop ~current_messages:updated_messages ~steps:(step :: steps) ~total_usage:new_total
                ~step_num:(step_num + 1))
        end
        else begin
          (* Final step *)
          let step : Generate_text_result.step =
            { text; reasoning; tool_calls; tool_results = []; finish_reason = fr; usage = step_usage }
          in
          Option.iter (fun f -> f step) on_step_finish;
          emit_event (Text_stream_part.Finish_step { finish_reason = fr; usage = step_usage });
          finish_stream ~finish_reason:fr ~usage:new_total ~all_steps:(List.rev (step :: steps))
        end
      end
    in
    Lwt.catch
      (fun () ->
        (* Execute pending tool approvals before starting the LLM step loop *)
        let%lwt start_messages, initial_steps =
          match pending_tool_approvals with
          | [] -> Lwt.return (initial_messages, [])
          | approvals ->
            emit_event Text_stream_part.Start_step;
            (* Emit tool input chunks so the frontend creates tool invocations *)
            List.iter
              (fun (ta : Generate_text_result.pending_tool_approval) ->
                emit_event
                  (Text_stream_part.Tool_call
                     { tool_call_id = ta.tool_call_id; tool_name = ta.tool_name; args = ta.args }))
              approvals;
            (* Denied before approved — matches upstream emit order *)
            approvals
            |> List.filter (fun (ta : Generate_text_result.pending_tool_approval) -> not ta.approved)
            |> List.iter (fun (ta : Generate_text_result.pending_tool_approval) ->
              emit_event (Text_stream_part.Tool_output_denied { tool_call_id = ta.tool_call_id }));
            let%lwt tool_results =
              Lwt_list.map_s
                (fun (ta : Generate_text_result.pending_tool_approval) ->
                  match ta.approved with
                  | false ->
                    Lwt.return
                      {
                        Generate_text_result.tool_call_id = ta.tool_call_id;
                        tool_name = ta.tool_name;
                        result = Core_tool.denied_result;
                        is_error = false;
                      }
                  | true ->
                    execute_and_emit { tool_call_id = ta.tool_call_id; tool_name = ta.tool_name; args = ta.args })
                approvals
            in
            let tool_calls =
              List.map
                (fun (ta : Generate_text_result.pending_tool_approval) ->
                  { Generate_text_result.tool_call_id = ta.tool_call_id; tool_name = ta.tool_name; args = ta.args })
                approvals
            in
            let step : Generate_text_result.step =
              {
                text = "";
                reasoning = "";
                tool_calls;
                tool_results;
                finish_reason = Ai_provider.Finish_reason.Tool_calls;
                usage = { input_tokens = 0; output_tokens = 0; total_tokens = Some 0 };
              }
            in
            Option.iter (fun f -> f step) on_step_finish;
            emit_event
              (Text_stream_part.Finish_step
                 {
                   finish_reason = Ai_provider.Finish_reason.Tool_calls;
                   usage = { input_tokens = 0; output_tokens = 0; total_tokens = Some 0 };
                 });
            (* Append tool results to messages for the next LLM call *)
            let tool_result_parts =
              List.map
                (fun (tr : Generate_text_result.tool_result) ->
                  {
                    Ai_provider.Prompt.tool_call_id = tr.tool_call_id;
                    tool_name = tr.tool_name;
                    result = tr.result;
                    is_error = tr.is_error;
                    content = [];
                    provider_options = Ai_provider.Provider_options.empty;
                  })
                tool_results
            in
            let updated_messages = initial_messages @ [ Ai_provider.Prompt.Tool { content = tool_result_parts } ] in
            Lwt.return (updated_messages, [ step ])
        in
        step_loop ~current_messages:start_messages ~steps:(List.rev initial_steps)
          ~total_usage:{ input_tokens = 0; output_tokens = 0; total_tokens = Some 0 }
          ~step_num:(1 + List.length initial_steps))
      (fun exn ->
        let msg = Printexc.to_string exn in
        push_full (Some (Text_stream_part.Error { error = msg }));
        push_full None;
        partial_output_push None;
        Lwt.wakeup_later_exn usage_resolver exn;
        Lwt.wakeup_later_exn finish_resolver exn;
        Lwt.wakeup_later_exn steps_resolver exn;
        Lwt.wakeup_later output_resolver None;
        Lwt.return_unit));
  {
    Stream_text_result.text_stream;
    full_stream;
    partial_output_stream;
    usage = usage_promise;
    finish_reason = finish_promise;
    steps = steps_promise;
    warnings = [];
    output = output_promise;
  }